Hugh DurrantWhyte
Australian Centre for Field Robotics
The University of Sydney NSW 2006
Australia
hugh@acfr.usyd.edu.au
September 27, 2002
Version 3.1
This material is BAE SYSTEMS Proprietary Information
This material is for ASCTA use only
c _Hugh DurrantWhyte 2001
1
BAE SYSTEMS Proprietary Information, ASCTA use only 2
Contents
1 Introduction 5
1.1 What is a Decentralised Data Fusion System ? . . . . . . . . . . . . . . . . 5
1.2 Course Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3 Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.4 Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2 Probabilistic Data Fusion 10
2.1 Probabilistic Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2 Probabilistic Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2.1 Bayes Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2.2 Data Fusion using Bayes Theorem . . . . . . . . . . . . . . . . . . . 16
2.2.3 Recursive Bayes Updating . . . . . . . . . . . . . . . . . . . . . . . 20
2.2.4 Bayesian Filtering and Tracking . . . . . . . . . . . . . . . . . . . . 24
2.2.5 Distributed Data Fusion with Bayes Theorem . . . . . . . . . . . . 28
2.2.6 Data Fusion with LogLikelihoods . . . . . . . . . . . . . . . . . . . 32
2.3 Information Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.3.1 Entropic Information . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.3.2 Conditional Entropy . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.3.3 Mutual Information . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.3.4 Fisher Information . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.3.5 The Relation between Shannon and Fisher Measures . . . . . . . . 44
3 MultiSensor Estimation 45
3.1 The Kalman Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.1.1 State and Sensor Models . . . . . . . . . . . . . . . . . . . . . . . . 46
3.1.2 The Kalman Filter Algorithm . . . . . . . . . . . . . . . . . . . . . 50
3.1.3 The Innovation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.1.4 Asynchronous, Delayed and Asequent Observations . . . . . . . . . 56
3.1.5 The Extended Kalman Filter . . . . . . . . . . . . . . . . . . . . . 61
3.1.6 The Covariance Intersect (CI) Algorithm . . . . . . . . . . . . . . . 64
3.2 The MultiSensor Kalman Filter . . . . . . . . . . . . . . . . . . . . . . . . 66
3.2.1 Observation Models . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.2.2 The GroupSensor Method . . . . . . . . . . . . . . . . . . . . . . . 71
3.2.3 The SequentialSensor Method . . . . . . . . . . . . . . . . . . . . . 75
3.2.4 The InverseCovariance Form . . . . . . . . . . . . . . . . . . . . . 77
3.2.5 TracktoTrack Fusion . . . . . . . . . . . . . . . . . . . . . . . . . 84
4 Decentralised Data Fusion Systems 88
4.1 Data Fusion Architectures . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
4.1.1 Hierarchical Data Fusion Architectures . . . . . . . . . . . . . . . . 89
4.1.2 Distributed Data Fusion Architectures . . . . . . . . . . . . . . . . 92
4.1.3 Decentralised Data Fusion Architectures . . . . . . . . . . . . . . . 93
BAE SYSTEMS Proprietary Information, ASCTA use only 3
4.2 Decentralised Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
4.2.1 The Information Filter . . . . . . . . . . . . . . . . . . . . . . . . . 96
4.2.2 The Information Filter and Bayes Theorem . . . . . . . . . . . . . . 101
4.2.3 The Information lter in MultiSensor Estimation . . . . . . . . . . 102
4.2.4 The Hierarchical Information Filter . . . . . . . . . . . . . . . . . . 105
4.2.5 The Decentralised Information Filter . . . . . . . . . . . . . . . . . 107
4.3 Decentralised MultiTarget Tracking . . . . . . . . . . . . . . . . . . . . . 110
4.3.1 Decentralised Data Association . . . . . . . . . . . . . . . . . . . . 110
4.3.2 Decentralised Identication and Bayes Theorem . . . . . . . . . . . 112
5 Communication in Decentralised Sensing Systems 114
5.1 Decentralised Communication Theory . . . . . . . . . . . . . . . . . . . . . 114
5.1.1 Communication Topologies . . . . . . . . . . . . . . . . . . . . . . . 114
5.1.2 Bayesian Communication in Sensor Networks . . . . . . . . . . . . 116
5.1.3 Identication of Redundant Information in Sensor Networks . . . . 118
5.1.4 Communication Channel Filters . . . . . . . . . . . . . . . . . . . . 120
5.1.5 Fully Connected and Broadcast Sensor Networks . . . . . . . . . . . 124
5.1.6 General Network Structures . . . . . . . . . . . . . . . . . . . . . . 125
5.2 Delays and Timing: Temporal Propagation of Information . . . . . . . . . 128
5.2.1 System Denitions . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
5.2.2 Forward Propagation of State . . . . . . . . . . . . . . . . . . . . . 129
5.2.3 Backward Propagation of State . . . . . . . . . . . . . . . . . . . . 129
5.2.4 Notation for Time Invariant Systems . . . . . . . . . . . . . . . . . 130
5.2.5 Denition of Estimates and Information States . . . . . . . . . . . . 131
5.2.6 Forward Time Propagation . . . . . . . . . . . . . . . . . . . . . . . 131
5.2.7 Backward Propagation . . . . . . . . . . . . . . . . . . . . . . . . . 132
5.2.8 Delayed Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
5.2.9 Information Observation and Update . . . . . . . . . . . . . . . . . 133
5.3 Delays in Channels and Burst Communication . . . . . . . . . . . . . . . . 134
5.3.1 Node Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
5.3.2 Channel Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . 135
5.4 The Communication Channel Algorithm . . . . . . . . . . . . . . . . . . . 136
5.4.1 Management of Communication . . . . . . . . . . . . . . . . . . . . 136
5.4.2 Structure of the Communication Algorithm . . . . . . . . . . . . . 138
5.4.3 Information Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
5.4.4 Computing the Common Information . . . . . . . . . . . . . . . . . 140
6 Management in Decentralised Systems 141
6.1 Bayesian Decision Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
6.1.1 Structure of the Problem . . . . . . . . . . . . . . . . . . . . . . . . 142
6.1.2 Utility or Loss . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
6.1.3 Expected Utility or Loss . . . . . . . . . . . . . . . . . . . . . . . . 143
6.1.4 Bayes Actions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
BAE SYSTEMS Proprietary Information, ASCTA use only 4
6.2 The Utility of Information . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
6.2.1 Information as Expected Utility . . . . . . . . . . . . . . . . . . . . 146
6.2.2 Information Filter Metrics . . . . . . . . . . . . . . . . . . . . . . . 147
6.2.3 Discrete State Metrics . . . . . . . . . . . . . . . . . . . . . . . . . 147
6.2.4 Composite Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
6.3 Decision Making with Multiple Information Sources . . . . . . . . . . . . . 148
6.3.1 The Super Bayesian . . . . . . . . . . . . . . . . . . . . . . . . . . 148
6.3.2 Multiple Bayesians . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
6.3.3 Practical Bargaining Solutions . . . . . . . . . . . . . . . . . . . . . 149
6.4 Decentralised Sensor Management . . . . . . . . . . . . . . . . . . . . . . . 151
6.4.1 Sensor to Target Assignment . . . . . . . . . . . . . . . . . . . . . . 151
6.4.2 Sensor Hando and Cueing . . . . . . . . . . . . . . . . . . . . . . 155
6.5 Communications Management . . . . . . . . . . . . . . . . . . . . . . . . . 157
6.6 Organisation and Control of Network Resource . . . . . . . . . . . . . . . . 159
7 Applications and Practical Considerations 163
7.1 Historical Projects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
7.1.1 SKIDS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
7.1.2 ISSS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
7.2 OxNav . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
7.2.1 Organisation and Management . . . . . . . . . . . . . . . . . . . . 166
7.2.2 OxNav System and Demonstrations . . . . . . . . . . . . . . . . . . 167
7.3 ANSER . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
7.3.1 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
7.3.2 Demonstrations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
7.3.3 Research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
7.4 Implementation of DDF Algorithms in ANSER . . . . . . . . . . . . . . . 175
7.4.1 Preprocessing and Coordinate Transformations . . . . . . . . . . . . 177
7.4.2 Data Association . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
7.4.3 Local Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
7.4.4 Channel Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
7.4.5 Channel Manager . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
7.4.6 Timing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
7.4.7 Picture Compilation . . . . . . . . . . . . . . . . . . . . . . . . . . 183
7.4.8 Terrain Navigation . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
7.4.9 The Communication Layer . . . . . . . . . . . . . . . . . . . . . . . 185
7.5 Large Scale Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . 186
BAE SYSTEMS Proprietary Information, ASCTA use only 5
1 Introduction
This course provides a practical introduction to decentralised data fusion (DDF) methods
and their application for engineers and scientists wishing to employ and develop DDF
systems. The course is mathematically advanced and assumes knowledge of estimation
theory and probabilistic model methods. The course emphasizes practical implementa
tions in ground and airborne sensor networks. A key feature of the course is the use
of laboratory sessions, based on Matlab, in which DDF methods are implemented and
evaluated by students. The intended outcomes of the course are to provide students with
the theoretical and practical skills necessary to design, implement and evaluate DDF
algorithms.
1.1 What is a Decentralised Data Fusion System ?
A decentralized data fusion system consists of a network of sensor nodes, each with its
own processing facility, which together do not require any central fusion or central com
munication facility. In such a system, fusion occurs locally at each node on the basis of
local observations and the information communicated from neighbouring nodes. At no
point is there a common place where fusion or global decisions are made.
A decentralised data fusion system is characterised by three constraints:
1. There is no single central fusion center; no one node should be central to the suc
cessful operation of the network.
2. There is no common communication facility; nodes cannot broadcast results and
communication must be kept on a strictly nodetonode basis.
3. Sensor nodes do not have any global knowledge of sensor network topology; nodes
should only know about connections in their own neighbourhood.
Figures 1, 2 and 3 show three possible realisations of a decentralised data fusion system
(these are discussed in more detail later in the course). The key point is that all these
systems have no central fusion center (unlike the decentralised systems often described
in the literature which are actually typically distributed or hierarchical).
The constraints imposed provide a number of important characteristics for decen
tralised data fusion systems:
Eliminating the central fusion center and any common communication facility en
sures that the system is scalable as there are no limits imposed by centralized
computational bottlenecks or lack of communication bandwidth.
Ensuring that no node is central and that no global knowledge of the network
topology is required for fusion means that the system can be made survivable to
the online loss (or addition) of sensing nodes and to dynamic changes in the network
structure.
BAE SYSTEMS Proprietary Information, ASCTA use only 6
As all fusion processes must take place locally at each sensor site and no global
knowledge of the network is required a priori, nodes can be constructed and pro
grammed in a modular fashion.
A decentralized system is characterised by being modular, scalable and survivable. To
gether, these give decentralised systems a substantial advantage over more traditional
sensing architectures in a range of defense and civilian data fusion tasks.
Sensor Node
Sensor
Fusion Processor
Communications Medium
Figure 1: A decentralised data fusion system implemented with a pointtopoint commu
nication architecture.
Decentralised data fusion methods capture many of the concepts embodied in network
centric system architectures. In particular, they immediately impose a requirement for
modularity, for scalability and robustness of the network, and for local intelligence at sen
sor nodes. However, unlike most current work in network centric systems, DDF methods
provide a mathematical and algorithmic foundation for the development of such systems.
1.2 Course Summary
This course consists of three main parts.
The rst part introduces the problem of data fusion and particularly distributed and
decentralised data fusion using probabilistic methods. Section 2 introduces Bayes theorem
BAE SYSTEMS Proprietary Information, ASCTA use only 7
Communications Medium
Sensor Node
Sensor
Fusion Processor
Figure 2: A decentralised data fusion system implemented with a broadcast, fully con
nected, communication architecture. Technically, a common communication facility vio
lates decentralised data fusion constraints. However a broadcast medium is often a good
model of real communication networks.
Sensor Payloads
Internal Communciation
External Communication
Figure 3: A decentralised data fusion system implemented with a hybrid, broadcast and
pointtopoint, communication architecture.
BAE SYSTEMS Proprietary Information, ASCTA use only 8
is the primary data fusion mechanism and describes how the algorithm can be distributed
and decentralised amongst a number of sensor nodes. The two basic problems in con
tinuous and discrete state estimation are described. Section 2 also introduces important
formal measures of Information. These are important in understanding the operation
of decentralised sensor networks and in further developing sensor and network manage
ment methods. Section 3 then turns attention to the problem of multisensor estimation.
Continuous state estimation (tracking) is probably the single most important problem in
data fusion as it aims to provide a single composite geometric picture of the environment.
Section 3 develops both the conventional multitarget multisensor Kalman lter and also
introduces the Information form of this problem. The Information form subsequently gives
rise to the decentralised Kalman lter and related decentralised tracking algorithms.
The second part of this course deals with the essential decentralised data fusion algo
rithms. Section 4 develops and describes the decentralised Kalman lter algorithm and its
application to multitarget multisensor tracking problems. The problem of decentralised
discrete estimation (classication for example) is also discussed in this context. The
operation of the algorithm, communication requirements, and structure of the resulting
sensor nodes are described. The modularity, scalability and survivability of the resulting
network algorithms is discussed and demonstrated. Section 5 focuses on the issue of com
munication in decentralised sensor networks. In particular communication algorithms for
broadcast, treeconnected, and arbitrary networks. Communication algorithms address
ing delayed and asequent data problems are developed. Issues of intermittent and burst
communication are also considered.
The third part of this course deals with some advanced principles in decentralised
data fusion with the objective of demonstrating how the basic algorithms are applied
to more complex and more realistic sensing problems. Section 5 considers the problem
of decentralised sensor and network management. This includes issues of sensorbased
decision making in a decentralised environment, (sensortotarget assignment and sensor
hando), as well as issues in communications and network management. Section 6 nally
looks at a number of practical implementations of the decentralised algorithms in both
past and current projects. The rst project describes the development of a fully modular
mobile robot implementing tracking, mapbuilding, platform control and sensor manage
ment algorithms. The second project is the current ANSER programme aiming to y
multipleUAVs cooperating in a fully decentralised form to build ground target pictures
and navigation maps. The potential future development of algorithms for very large scale
decentralised ground sensor networks is also discussed.
1.3 Bibliography
There is relatively little material on decentralised data fusion methods; indeed these course
notes are probably the most comprehensive assembly of such information. A comprehen
sive bibliography of the citations used in this work is included at the end of these course
notes. Here we comment on some general sources:
The general data fusion is well covered in the recent book by Blackman and Popoli
BAE SYSTEMS Proprietary Information, ASCTA use only 9
[11]. It covers both level 12 multitarget tracking and identication problems as well as
level 34 methods in situation assessment and sensor management. Notably, it covers a
number of current military systems in some detail and gives develops a number of specic
examples of multisensor systems. The books by Barshalom are also useful sources [5, 3, 4].
There are other notable data fusion books ([63] for example), although these tend to be
largely oriented around block diagrams.
There are two books on decentralised data fusion published by some of my previous
Ph.D. students. The book by James Manyika [36] focuses on problems in sensor manage
ment but also captures some important aspects in the implementation of decentralised
systems. The book by Arthur Mutambara [42] is mainly concerned with the problem of
decentralised estimation and control. Both books reect the stateoftheart in 1995.
A useful source of detailed information are the theses of my previous Ph.D. students.
I can often obtain or make available copies of these if required. The theses are generally
listed in the citations at the end of these course notes. In addition there are a signicant
number of papers published in the open literature over the past decade.
Finally there is project documentation, particularly for the current ANSER project.
This is usually the most detailed and most current information available. However, de
tailed project documentation is not normally widely distributed: requests should be made
through the programme manager at BAE Systems.
1.4 Acknowledgements
The work described in this course is the result of over a decade of research by myself, my
students and research sta, and BAE Systems.
A large number of outstanding graduate students have been involved in this re
search: Bobby Rao, Stewart Grime, Tim Berg, James Manyika, Mike Stevens, Peter Ho,
Tom Burke, Mariano Fernandez, Simukai Utete, Arthur Mutambara, Simon Julier, Jef
frey Uhlmann, Rob Deaves and Eric Nettleton. The course draws substantially on the
work described in their theses.
The research would not have been possible without the considerably and continuous
support of Phil Greenway at BAE Systems Sowerby Research Centre. A number of sta at
Sowerby have also made substantial contributions to the work described here, including:
Rob Deaves, Dave Nicholson, Julia Suttclie, and Rob Dawkins.
BAE SYSTEMS Proprietary Information, ASCTA use only 10
2 Probabilistic Data Fusion
Uncertainty lies at the heart of all descriptions of the sensing and data fusion process.
An explicit measure of this uncertainty must be provided to enable sensory information
to be fused in an ecient and predictable manner. Although there are many methods
of representing uncertainty, almost all of the theoretical developments in this course are
based on the use of probabilistic models. There is a huge wealth of experience and methods
associated with the development of probabilistic models of environments and information.
Probabilistic models provide a powerful and consistent means of describing uncertainty
in a broad range of situations and leads naturally into ideas of information fusion and
decision making.
This section begins by briey introducing the essential elements of probabilistic model
ing: probability densities, conditional densities, and Bayes theorem. With these elements,
the basic data fusion problem is described and it is shown how information can be com
bined in a probabilistic manner. From this, the construction of various data fusion archi
tectures is described. It is demonstrated how the mathematics required for data fusion in
these architectures can be developed and implemented directly from Bayes theorem. This
methodology underlies many of the decentralised data fusion algorithms developed in the
remainder of the this course. The idea of information and entropy are then introduced.
Entropy measures are a natural way of describing the ow of information in dierent
data fusion architectures. It is demonstrated how these information measures can be used
to predict, control and manage the data fusion process. These methods, together with
the information form of the Kalman lter are the underlying basis for the decentralised
data fusion methods developed in this course.
2.1 Probabilistic Models
In the following, familiarity with essential probability theory is assumed, and some simple
notation and rules to be used throughout this course are introduced. A probability density
function (pdf ) P
y
() is dened on a random variable y, generally written as P
y
(y) or simply
P(y) when the dependent variable is obvious. The random variable may be a scalar or
vector quantity, and may be either discrete or continuous in measure.
The pdf is considered as a (probabilistic) model of the quantity y; observation or state.
The pdf P(y) is considered valid if;
1. It is positive; P(y) 0 for all y, and
2. It sums (integrates) to a total probability of 1;
_
y
P(y)dy = 1.
The joint distribution P
xy
(x, y) is dened in a similar manner.
BAE SYSTEMS Proprietary Information, ASCTA use only 11
Integrating the pdf P
xy
(x, y) over the variable x gives the marginal pdf P
y
(y) as
P
y
(y) =
_
x
P
xy
(x, y)dx, (1)
and similarly integrating over y gives the marginal pdf P
x
(x). The joint pdf over n
variables, P(x
1
, , x
n
), may also be dened with analogous properties to the joint pdf
of two variables.
The conditional pdf P(x [ y) is dened by
P(x [ y)
=
P(x, y)
P(y)
, (2)
and has the usual properties of a pdf with x the dependent variable given that y takes on
specic xed values. The conditional pdf P(y [ x) is similarly dened.
The chainrule of conditional distributions can be used to expand a joint pdf in terms
of conditional and marginal distributions. From Equation 2,
P(x, y) = P(x [ y)P(y). (3)
The chainrule can be extended to any number of variables in the following form
P(x
1
, , x
n
) = P(x
1
[ x
2
, x
n
) P(x
n1
[ x
n
)P(x
n
), (4)
where the expansion may be taken in any convenient order. Substitution of Equation 3
into Equation 1 gives an expression for the marginal distribution of one variable in terms
of the marginal distribution of a second variable as
P
y
(y) =
_
x
P
xy
(y [ x)P
x
(x)dx. (5)
This important equation is known as the total probability theorem. It states that the
total probability in a state y can be obtained by considering the ways in which y can
occur given that the state x takes a specic value (this is encoded in P
xy
(y [ x)), weighted
by the probability that each of these values of x is true (encoded in P
x
(x)).
If it happens that knowledge of the value of y does not give us any more information
about the value of x then x and y are said to be independent as
P(x [ y) = P(x). (6)
With Equation 6 substituted into Equation 3
P(x, y) = P(x)P(y). (7)
A weaker form of independence can be dened through the important idea of conditional
independence. Given three random variables x, y and z, the conditional distribution of
BAE SYSTEMS Proprietary Information, ASCTA use only 12
x given both y and z is dened as P(x [ yz). If knowledge of the value of z makes the
value of x independent of the value of y then
P(x [ y, z) = P(x [ z). (8)
This may be the case for example if z indirectly contains all the information contributed
by y to the value of x. Conditional independence can be exploited in a number of dierent
ways. In particular, applying the chainrule to the joint probability density function on
three random variables x, y, and z
P(x, y, z) = P(x, y [ z)P(z)
= P(x [ y, z)P(y [ z)P(z),
(9)
together with the conditional independence result of Equation 8 the intuitive result
P(x, y [ z) = P(x [ z)P(y [ z), (10)
is obtained. That is, if x is independent of y given knowledge of z then the joint probability
density function of x and y conditioned on z is simply the product of the marginal
distributions of x and y each conditioned on z, analogously to Equation 7.
The idea of conditional independence underlies many data fusion algorithms. Consider
the state of a system x and two observations of this state z
1
and z
2
. It should be clear
that the two observations are not independent,
P(z
1
, z
2
) ,= P(z
1
)P(z
2
),
as they must both depend on the common state x. Indeed, if the two observations were
independent (were unrelated to each other), there would be little point fusing the infor
mation they contain ! Conversely, it is quite reasonable to assume that the only thing the
two observations have in common is the underlying state, and so the observations are in
dependent once the state is known; that is, the observations are conditionally independent
given the state as
P(z
1
, z
2
[ x) = P(z
1
[ x)P(z
2
[ x).
Indeed, for the purposes of data fusion, this would not be a bad denition of the state;
simply what the two information sources have in common.
2.2 Probabilistic Methods
2.2.1 Bayes Theorem
Bayes theorem is arguably the most important result in the study of probabilistic models.
Consider two random variables x and z on which is dened a joint probability density
function P(x, z). The chainrule of conditional probabilities can be used to expand this
density function in two ways
P(x, z) = P(x [ z)P(z)
= P(z [ x)P(x). (11)
BAE SYSTEMS Proprietary Information, ASCTA use only 13
Rearranging in terms of one of the conditional densities, Bayes theorem is obtained
P(x [ z) =
P(z [ x)P(x)
P(z)
. (12)
The value of this result lies in the interpretation of the probability density functions
P(x [ z), P(z [ x), and P(x). Suppose it is necessary to determine the various likelihoods
of dierent values of an unknown state of nature x A. There may be prior beliefs
about what values of x might be expect, encoded in the form of relative likelihoods in
the prior probability density function P(x). To obtain more information about
the state x an observation z Z is made. The observations made are modeled as a
conditional probability density function P(z [ x) which describes, for each xed state of
nature x A, the likelihood that the observation z Z will be made; the probability
of z given x. The new likelihoods associated with the state of nature x must now be
computed from the original prior information and the information gained by observation.
This is encoded in the posterior distribution P(x [ z) which describes the likelihoods
associated with x given the observation z. The marginal distribution P(z) simply serves
to normalize the posterior. The value of Bayes theorem is now clear, it provides a direct
means of combining observed information with prior beliefs about the state of the world.
Unsurprisingly, Bayes theorem lies at the heart of many data fusion algorithms.
The conditional distribution P(z [ x) serves the role of a sensor model. This distribu
tion can be thought of in two ways. First, in building a sensor model, the distribution
is constructed by xing the value
1
of x = x and then asking what pdf in the variable z
results. Thus, in this case, P(z [ x) is considered as a distribution on z. For example, if
we know the true range to a target (x), then P(z [ x) is the distribution on the actual ob
servation of this range. Conversely, once the sensor model exists, observations (numbers,
not distributions) are made and z = z is xed. From this however, we want to infer the
state x. Thus the distribution P(z [ x) is now considered as a distribution in x. In this
latter, case, the distribution is known as the Likelihood Function and the dependence
on x is made clear by writing (x) = P(z [ x).
In a practical implementation of Equation 12, P(z [ x) is constructed as a function of
both variables (or a matrix in discrete form). For each xed value of x, a distribution in
z is dened. Therefore as x varies, a family of distributions in z is created. The following
two examples, one in continuous variables and one in discrete variables, makes these ideas
clear.
Example 1
Consider a continuous valued state x, the range to target for example, and an ob
servation z of this state. A commonly used model for such an observation is where the
observation made of true state is Normally (Gaussian) distributed with mean x and a
variance
2
z
as
P(z [ x) =
1
2
2
z
exp
_
1
2
(z x)
2
2
z
_
. (13)
1
Random variables are denoted by a bold face font, specic values taken by the random variable are
denoted by normal fonts.
BAE SYSTEMS Proprietary Information, ASCTA use only 14
It should be clear that this is a simple function of both z and x. If we know the true
value of the state, x = x, then the distribution is a function of z only; describing the
probability of observing a particular value of range (Normally distributed around the true
range x with variance
2
z
). Conversely, if we make a specic observation, z = z, then
the distribution is a function of x only; describing the probability of the true range value
(Normally distributed around the range observation z with variance
2
z
). In this case, the
distribution is the Likelihood Function.
Now assume that we have some prior belief about the true state x encoded in a Gaussian
prior as
P(x) =
1
2
2
x
exp
_
1
2
(x x
p
)
2
2
x
_
.
Note, this is a function of a single variable x (with x
p
xed). Bayes theorem can be
directly applied to combine this prior information with information from a sensor, modeled
by 13. First, an observation (technically, an experimental realisation), z, is made and
instantiated in Equation 13. Then the prior and sensor model are multiplied together to
produce a posterior distribution (which is a function of x only, and is sometimes referred
to as the posterior likelihood) as
P(x [ z) = C
1
2
2
z
exp
_
1
2
(x z)
2
2
z
_
.
1
2
2
x
exp
_
1
2
(x x
p
)
2
2
x
_
(14)
=
1
2
2
exp
_
1
2
(x x)
2
2
_
(15)
where C is a constant independent of x chosen to ensure that the posterior is appropriately
normalized, and x and
2
are given by
x =
2
x
2
x
+
2
z
z +
2
z
2
x
+
2
z
x
p
,
and
2
=
2
z
2
x
2
z
+
2
x
=
_
1
2
z
+
1
2
x
_
1
Thus the posterior is also Gaussian with a mean that is the weighted average of the means
for the original prior and likelihood, and with a variance equal to the parallel combination
of the original variances.
Example 2
Consider a simple example of the application of Bayes theorem to estimating a discrete
parameter on the basis of one observation and some prior information. The environment
of interest is modeled by a single state x which can take on one of three values:
x
1
: x is a type 1 target.
BAE SYSTEMS Proprietary Information, ASCTA use only 15
x
2
: x is a type 2 target.
x
3
: No visible target.
A single sensor observes x and returns three possible values:
z
1
: Observation of a type 1 target.
z
2
: Observation of a type 2 target.
z
3
: No target observed.
The sensor model is described by the likelihood matrix P
1
(z [ x):
z
1
z
2
z
3
x
1
0.45 0.45 0.1
x
2
0.45 0.45 0.1
x
3
0.1 0.1 0.8
Note, that this likelihood matrix is a function of both x and z. For a xed value of the
true state, it describes the probability of a particular observation being made (the rows of
the matrix). When a specic observation is made, it describes a probability distribution
over the values of true state (the columns) and is then the Likelihood Function (x).
The posterior distribution of the true state x after making an observation z = z
i
is
given by
P(x [ z
i
) = P
1
(z
i
[ x)P(x)
where is simply a normalizing constant set by requiring the sum, over x, of posterior
probabilities to be equal to 1.
In the rst instance we will assume that we do not have any prior information about
the possible likelihood of target types 1 and 2, and so we set the prior probability vector
to P(x) = (0.333, 0.333, 0.333). If we now observe z = z
1
, then clearly the posterior
distribution will be given by P(x [ z
1
) = (0.45, 0.45, 0.1) (i.e. the rst column of the
likelihood matrix above; the likelihood function given z
1
has been observed).
If now we subsequently use this posterior distribution as the prior for a second obser
vation P(x) = (0.45, 0.45, 0.1), and we again make the observation z = z
1
, then the new
posterior distribution will be given by
P(x [ z
1
) = P
1
(z
1
[ x)P(x)
= (0.45, 0.45, 0.1) (0.45, 0.45, 0.1)
= (0.488, 0.488, 0.024).
(where the notation denotes an elementwise product).
Notice that the result of this integration process is to increase the probability in both
type 1 and type 2 targets at the expense of the notarget hypothesis. Clearly, although
this sensor is good at detecting targets, it is not good at distinguishing between targets of
dierent types.
BAE SYSTEMS Proprietary Information, ASCTA use only 16
2.2.2 Data Fusion using Bayes Theorem
It is possible to apply Bayes theorem directly to the integration of observations from
several dierent sources. Consider the set of observations
Z
n
= z
1
Z
1
, , z
n
Z
n
.
It is desired to use this information to construct a posterior distribution P(x [ Z
n
)
describing the relative likelihoods of the various values of the state of interest x A
given the information obtained. In principle, Bayes theorem can be directly employed to
compute this distribution function from
P(x [ Z
n
) =
P(Z
n
[ x)P(x)
P(Z
n
)
=
P(z
1
, , z
n
[ x)P(x)
P(z
1
, , z
n
)
.
(16)
In practice it would be dicult to do this because it requires that the joint distribution
P(z
1
, , z
n
[ x) is known completely; that is, the joint distribution of all possible combi
nations of observations conditioned on the underlying state
2
. However, it is usually quite
reasonable to assume that given the true state x A, the information obtained from the
i
th
information source is independent of the information obtained from other sources. The
validity of this assumption is discussed below. With this assumption, Equation 8 implies
that
P(z
i
[ x, z
1
, , z
i1
, z
i+1
, , z
n
) = P(z
i
[ x), (17)
and from Equation 10 this gives
P(z
1
, , z
n
[ x) = P(z
1
[ x) P(z
n
[ x) =
n
i=1
P(z
i
[ x). (18)
Substituting this back into Equation 16 gives
P(x [ Z
n
) = [P(Z
n
)]
1
P(x)
n
i=1
P(z
i
[ x). (19)
Thus the updated likelihoods in the state, the posterior distribution on x, is simply propor
tional to the product of prior likelihood and individual likelihoods from each information
source. The marginal distribution P(Z
n
) simply acts as a normalising constant. Equa
tion 19 provides a simple and direct mechanism for computing the relative likelihood in
dierent values of a state from any number of observations or other pieces of information.
Equation 19 is known as the independent likelihood pool [10]. In practice, the condi
tional probabilities P(z
i
[ x) are stored a priori as functions of both z
i
and x. When an
observation sequence Z
n
= z
1
, z
2
, , z
n
is made, the observed values are instantiated
2
In Example 2 above, this would require the construction of likelihood matrix of size m
n
where m is
the number of possible outcomes for each observation and where n is the number of observations made.
BAE SYSTEMS Proprietary Information, ASCTA use only 17
in this probability distribution and likelihood functions
i
(x) are constructed, which are
functions only of the unknown state x. The product of these likelihood functions with
the prior information P(x), appropriately normalised, provides a posterior distribution
P(x [ Z
n
), which is a function of x only for a specic observation sequence z
1
, z
2
, , z
n
.
Figure 4 shows the structure of the independent likelihood pool in a centralised architec
ture.
z
1
z
n
z
i
Central
Processor
1
(x)
i
(x)
n
(x)
P(x)
P(z
n
x)
P(xZ
n
)=C P( x)
i
(x)
n
i=1
P(z
i
x) P(z
1
x)
X
Figure 4: The centralised implementation of the independent likelihood pool as a method
for combining information from a number of sources. The central processor maintains a
model of each sensor i in terms of a conditional probability distribution P
i
(z
i
[ x), together
with any prior probabilistic knowledge P(x). On arrival of a measurement set Z
n
, each
sensor model is instantiated with the associated observation to form a likelihood
i
(x).
The normalised product of these yields the posterior distribution P(x [ Z
n
).
The eectiveness of Equation 19 relies crucially on the assumption that the information
obtained from dierent information sources is independent when conditioned on the true
underlying state of the world; this is dened in Equation 17. It would be right to question
if this assumption is reasonable. It is clearly unreasonable to state that the information
obtained is unconditionally independent;
P(z
1
, , z
n
) ,= P(z
1
) P(z
n
), (20)
because each piece of information depends on a common underlying state x A. If the
information obtained were independent of this state, and therefore unconditionally inde
pendent of other information sources, there would be little value in using it to improve
BAE SYSTEMS Proprietary Information, ASCTA use only 18
knowledge of the state. It is precisely because the information obtained is unconditionally
dependent on the underlying state that it has value as an information source. Conversely,
it is generally quite reasonable to assume that the underlying state is the only thing in
common between information sources and so once the state has been specied it is corre
spondingly reasonable to assume that the information gathered is conditionally indepen
dent given this state. There are sometimes exceptions to this general rule, particularly
when the action of sensing has a nontrivial eect on the environment.
Example 3
Consider again Example 2 of the discrete observation of target type. A second sensor is
obtained which makes the same three observations as the rst sensor, but whose likelihood
matrix P
2
(z
2
[ x) is described by
z
1
z
2
z
3
x
1
0.45 0.1 0.45
x
2
0.1 0.45 0.45
x
3
0.45 0.45 0.1
Whereas the rst sensor was good at detecting targets but not at distinguishing between
dierent target types, this second sensor has poor overall detection probabilities but good
target discrimination capabilities. So for example, with a uniform prior, if we observe
z = z
1
with this second sensor, the posterior distribution on possible true states will be
given by P(x [ z
1
) = (0.45, 0.1, 0.45) (i.e the rst column of the likelihood matrix) .
It clearly makes sense to combine the information from both sensors to provide a
system with both good detection and good discrimination capabilities. From Equation 19,
the product of the two likelihood functions gives us an overall likelihood function for the
combined system as P
12
(z
1
, z
2
[ x) = P
1
(z
1
[ x)P
2
(z
2
[ x). Thus if we observe z
1
= z
1
using the rst sensor, and z
2
= z
1
with the second sensor (assuming a uniform prior),
then the posterior likelihood in x is given by
P(x [ z
1
, z
1
) = P
12
(z
1
, z
1
[ x)
= P
1
(z
1
[ x)P
2
(z
1
[ x)
= (0.45, 0.45, 0.1) (0.45, 0.1, 0.45)
= (0.6924, 0.1538, 0.1538)
Comparing this to taking two observations of z
1
with sensor 1 (in which the resulting
posterior was (0.488, 0.488, 0.024)) it can be seen that sensor 2 adds substantial target
discrimination power at the cost of a slight loss of detection performance for the same
number of observations.
Repeating this calculation for each z
1
, z
2
observation pair, results in the combined
BAE SYSTEMS Proprietary Information, ASCTA use only 19
likelihood matrix
z
1
= z
1
z
2
= z
1
z
2
z
3
x
1
0.6924 0.1538 0.4880
x
2
0.1538 0.6924 0.4880
x
3
0.1538 0.1538 0.0240
z
1
= z
2
z
2
= z
1
z
2
z
3
x
1
0.6924 0.1538 0.4880
x
2
0.1538 0.6924 0.4880
x
3
0.1538 0.1538 0.0240
z
1
= z
3
z
2
= z
1
z
2
z
3
x
1
0.1084 0.0241 0.2647
x
2
0.0241 0.1084 0.2647
x
3
0.8675 0.8675 0.4706
The combined sensor provides substantial improvements in overall system performance
3
.
If for example we observe target 1 with the rst sensor (the array block z
1
= z
1
) and again
observe target 1 with the second sensor (the rst column of this block), then the posterior
distribution in the three hypotheses is
P(x [ z
1
, z
2
) = (0.692, 0.154, 0.154),
and so target 1 is clearly the most probable target. If however, we observe a type 2
target with the second sensor after having observed a type 1 target with the rst sensor,
a similar calculation gives the posterior as (0.154, 0.692, 0.154), that is target type 2 has
high probability. This is because although sensor 1 observed a type 1 target, the likelihood
function for sensor 1 tells us that it is poor at distinguishing between target types and so
sensor 2 information is used for this purpose. If now we observe no target with sensor 2,
3
Note that summing over any column still come to 1. In practical implementations, it is often sucient
to encode relative likelihoods of dierent events and to normalize only when computing the posterior
distribution.
BAE SYSTEMS Proprietary Information, ASCTA use only 20
having detected target type 1 with the rst sensor, the posterior given both observations
is given by (0.488, 0.488, 0.024). That is we still believe that there is a target (because
we know sensor 1 is much better at target detection than sensor 2), but we still have no
idea which of target 1 or 2 it is as sensor 2 has been unable to make a valid detection.
The analysis for sensor 1 detecting target 2 is identical to that for detection of target 1.
Finally, if sensor 1 gets no detection, but sensor 2 detects target type 1, then the posterior
likelihood is given by (0.108, 0.024, 0.868). That is we still believe there is no target because
we know sensor 1 is better at providing this information (and perversely, sensor 2 conrms
this even though it has detected target type 1).
Practically, the joint likelihood matrix is never constructed (it is easy to see why here,
with n = 3 sensors, and m = 3 possible observations and k = 3 possible outcomes, the
dimension of the joint likelihood matrix has k m
n
= 27 entries.) Rather, the likelihood
matrix is constructed for each sensor and these are only combined when instantiated with
an observation. Storage then reduces to n arrays of dimension k m, at the cost of a
k dimensional vector multiply of the instantiated likelihood functions. This is clearly a
major saving in storage and complexity and underlines the importance of the conditional
independence assumption to reduction in computational complexity.
2.2.3 Recursive Bayes Updating
The integration of information using Equation 19 would, in principle, require that all past
information is remembered and, on arrival of new information in the form P(z
k
[ x), that
the total likelihood be recomputed based on all information gathered up to this time.
However, Bayes theorem, in the form of Equation 19, also lends itself to the incremental
or recursive addition of new information in determining a revised posterior distribution
on the state. With Z
k
= z
k
, Z
k1
P(x, Z
k
) = P(x [ Z
k
)P(Z
k
) (21)
= P(z
k
, Z
k1
[ x)P(x)
= P(z
k
[ x)P(Z
k1
[ x)P(x), (22)
where it is assumed conditional independence of the observation sequence. Equating both
sides of this expansion gives
P(x [ Z
k
)P(Z
k
) = P(z
k
[ x)P(Z
k1
[ x)P(x) (23)
= P(z
k
[ x)P(x [ Z
k1
)P(Z
k1
). (24)
Noting that P(Z
k
)/P(Z
k1
) = P(z
k
[ Z
k1
) and rearranging gives
P(x [ Z
k
) =
P(z
k
[ x)P(x [ Z
k1
)
P(z
k
[ Z
k1
)
. (25)
The advantage of Equation 25 is that we need compute and store only the posterior
likelihood P(x [ Z
k1
) which contains a complete summary of all past information. When
BAE SYSTEMS Proprietary Information, ASCTA use only 21
the next piece of information P(z
k
[ x) arrives, the previous posterior takes on the role
of the current prior and the product of the two becomes, when normalised, the new
posterior. Equation 25 thus provides a signicant improvement in computational and
memory requirements over Equation 19.
Example 4
An important example of the recursive application of Bayes theorem is in the calcu
lation of the posterior distribution of a scalar x under the assumption that the likelihood
function for the observations given the true state is Gaussian with known variance
2
;
P(z
k
[ x) =
1
2
2
exp
_
1
2
(z
k
x)
2
2
_
.
If we assume that the posterior distribution in x after taking the rst k 1 observations
is also Gaussian with mean x
k1
and variance
2
k1
,
P(x [ Z
k1
) =
1
2
2
k1
exp
_
1
2
(x
k1
x)
2
2
k1
_
.
then the posterior distribution in x after the rst k observations is given by
P(x [ Z
k
) = K
1
2
2
exp
_
1
2
(z
k
x)
2
2
_
.
1
2
2
k1
exp
_
1
2
(x
k1
x)
2
2
k1
_
(26)
=
1
2
2
k
exp
_
1
2
(x
k
x)
2
2
k
_
(27)
where K is a constant independent of x chosen to ensure that the posterior is appropriately
normalized, and x
k
and
2
k
are given by
x
k
=
2
k1
2
k1
+
2
z
k
+
2
2
k1
+
2
x
k1
, (28)
and
2
k
=
2
2
k1
2
+
2
k1
(29)
The most important point to note about the posterior is that it too is a Gaussian; the
product of two Gaussian distributions is itself Gaussian. Distributions that have this sym
metry property are known as conjugate distributions. Given this property, it is clearly
not necessary to go through the process of multiplying distributions together as it is su
cient to simply compute the new mean and variance recursively from Equations 28 and 29
as these completely characterize the associated Gaussian distribution. With these simple
recursion equations any number of observations may be fused in a simple manner.
BAE SYSTEMS Proprietary Information, ASCTA use only 22
0
10
20
30
40
50
0
10
20
30
40
50
0
2
4
6
x 10
3
X Range (km)
Prior Location Density
Y Range (km)
(a)
0
10
20
30
40
50
0
10
20
30
40
50
0
0.01
0.02
0.03
0.04
X Range (km)
Location likelihood from Sensor 1
Y Range (km)
(b)
0
10
20
30
40
50
0
10
20
30
40
50
0
0.02
0.04
0.06
X Range (km)
Posterior location density after one observation from sensor 1
Y Range (km)
(c)
0
10
20
30
40
50
0
10
20
30
40
50
0
0.02
0.04
0.06
0.08
0.1
X Range (km)
Posterior location density after two observations from sensor 1
Y Range (km)
(d)
0
10
20
30
40
50
0
10
20
30
40
50
0
0.01
0.02
0.03
0.04
X Range (km)
Location likelihood from Sensor 2
Y Range (km)
(e)
0
10
20
30
40
50
0
10
20
30
40
50
0
0.2
0.4
0.6
0.8
1
X Range (km)
Posterior location density following update from sensor 2
Y Range (km)
(f)
Figure 5: Generalised Bayes Theorem. The gures show plots of twodimensional distri
bution functions dened on a grid of x and y points: (a) Prior distribution; (b) likelihood
function for rst sensor; (c) posterior after one application of Bayes Theorem; (d) posterior
after two applications; (e) likelihood function for second sensor; (f) nal posterior.
BAE SYSTEMS Proprietary Information, ASCTA use only 23
Example 5
Quite general prior and likelihood distributions can be handled by direct application of
Bayes Theorem. Consider the problem in which we are required to determine the location
of a target in a dened area. Figure 5(a) shows a general prior probability distribution de
ned on the search area. The distribution is simply dened as a probability value P(x
i
, y
j
)
on a grid point at location x
i
, y
j
, of area dx
i
, dy
j
. The only constraints placed on this
distribution are that
P(x
i
, y
j
) > 0, x
i
, y
j
and that
j
P(x
i
, x
j
)dx
i
dy
i
= 1
(which can easily be satised by appropriate normalisation.)
A sensor (sensor 1) now takes observations of the target from a sensor located at
x = 15, y = 0km. The likelihood function generated from this sensor following an ob
servation z
1
is shown in Figure 5(b). This likelihood P
1
(z
1
[ x
i
, y
j
) again consists of a
general location probability dened on the x
i
y
j
grid. The likelihood shows that the bearing
resolution of the sensor is high, whereas it has almost no range accuracy (the likelihood is
long and thin with probability mass concentrated on a line running from sensor to target).
The posterior distribution having made this rst observation is shown in Figure 5(c) and
is computed from the pointwise product of prior and likelihood,
P(x
i
, y
j
[ z
1
) = P
1
(z
1
[ x
i
, y
j
) P(x
i
, y
j
),
where is simply a normalising constant. It can be seen that the distribution dening
target location is now approximately restrained to a line along the detected bearing. Figure
5(d) shows the posterior P(x
i
, y
j
[ z
1
, z
2
) following a second observation z
2
by the same
sensor. This is again computed by pointwise multiplication of the likelihood P
1
(z
2
[ x
i
, y
j
)
with the new prior (the posterior from the previous observation P(x
i
, y
j
[ z
1
)). It can be
seen that there is little improvement in location density following this second observation;
this is to be expected as there is still little range data available.
A second sensor (sensor 2) now takes observations of the target from a location x =
50, y = 20. Figure 5(e) shows the target likelihood P
2
(z
3
[ x
i
, y
j
) following an observation
z
3
by this sensor. It can be seen that this sensor (like sensor 1), has high bearing resolution,
but almost no range resolution. However, because the sensor is located at a dierent site,
we would expect that the combination of bearing information from the two sensors would
provide accurate location data. Indeed, following pointwise multiplication of the second
sensor likelihood with the new prior (the posterior P(x
i
, y
j
[ z
1
, z
2
) from the previous two
observations of sensor 1), we obtain the posterior P(x
i
, y
j
[ z
1
, z
2
, z
3
) shown in Figure
5(f ) which shows all probability mass highly concentrated around a single target location.
BAE SYSTEMS Proprietary Information, ASCTA use only 24
2.2.4 Bayesian Filtering and Tracking
The general ltering and tracking problem can be formulated in Bayesian form. This is
signicant because it provides a common representation for a range of discrete and contin
uous data fusion problems without recourse to specic target or observation models. This
general representation forms a base class from which specic data fusion, tracking and
identication tasks can be derived. Distributed data fusion algorithms for specic network
architectures will subsequently be formulated with respect to this general representation.
From these, specic decentralised data fusion algorithms can be derived.
Consider a state x
t
which is generally taken to be a function of time k. This may, for
example, describe a target to be tracked or the location of a platform for which navigation
data is required. For convenience and without loss of generality, time is dened at discrete
(asynchronous) times t
k
i=1
P(z
i
[ X
k
) =
k
i=1
P(z
i
[ x
i
). (32)
The state transition model can be described in terms of a probability distribution in
the form
P(x
k
[ x
k1
, u
k
). (33)
That is, the state transition may reasonably be assumed to be a Markov process in which
the next state x
k
depends only on the immediately proceeding state x
k1
and the applied
control u
k
, and is independent of both the observations and preceding states. With these
denitions and models, Bayes Theorem may be employed to dene a recursive solution to
Equation 30.
To derive a recursive update rule for the posterior, the chain rule of conditional prob
ability is employed to expand the joint distribution of the state and observation in terms
of the state
P(x
k
, z
k
[ Z
k1
, U
k
, x
0
) = P(x
k
[ z
k
, Z
k1
, U
k
, x
0
)P(z
k
[ Z
k1
, U
k
, x
0
)
= P(x
k
[ Z
k
, U
k
, x
0
)P(z
k
[ Z
k1
U
k
), (34)
and then in terms of the observation
P(x
k
, z
k
[ Z
k1
, U
k
, x
0
) = P(z
k
[ x
k
, Z
k1
, U
k
, x
0
)P(x
k
[ Z
k1
, U
k
, x
0
)
= P(z
k
[ x
k
)P(x
k
[ Z
k1
, U
k
, x
0
) (35)
where the last equality employs the assumptions established for the sensor model in
Equation 31.
Equating Equations 34 and 35 and rearranging gives
P(x
k
[ Z
k
, U
k
, x
0
) =
P(z
k
[ x
k
)P(x
k
[ Z
k1
, U
k
, x
0
)
P(z
k
[ Z
k1
, U
k
)
. (36)
The denominator in Equation 36 is independent of the state and can therefore be set to
some normalising constant K.
Equation 36 is the observation update step for the full Bayes data fusion algorithm. A
graphical description of this equation is shown in Figure 6.
The total probability theorem can be used to rewrite the second term in the numerator
of Equation 36 in terms of the state transition model and the joint posterior from timestep
k 1 as
P(x
k
[ Z
k1
, U
k
, x
0
) =
_
P(x
k
, x
k1
[ Z
k1
, U
k
x
0
)dx
k1
=
_
P(x
k
[ x
k1
, Z
k1
, U
k
, x
0
)P(x
k1
[ Z
k1
, U
k
, x
0
)dx
k1
=
_
P(x
k
[ x
k1
, u
k
)P(x
k1
[ Z
k1
, U
k1
, x
0
)dx
k1
(37)
BAE SYSTEMS Proprietary Information, ASCTA use only 26
0
10
20
30
40
50
0
5
10
15
0
0.2
0.4
0.6
0.8
1
1.2
X
Z
P(z
k
x
k
= x
1
)
P(z
k
x
k
= x
2
)
P(z
k
= x
1
x
k
)
P( x
k
)
Figure 6: Observation update for the full Bayes lter. Prior to observation, an observation
model in the form of the conditional density P(z
k
[ x
k
) is established. For a xed value of
x
k
, equal to x
1
or x
2
for example, a density function P(z
k
[ x
k
= x
1
) or P(z
k
[ x
k
= x
2
)
is dened describing the likelihood of making the observation z
k
. Together the density
P(z
k
[ x
k
) is then a function of both z
k
and x
k
. This conditional density then denes
the observation model. Now, in operation, a specic observation z
k
= x
1
is made and
the resulting distribution P(z
k
= x
1
[ x
k
) denes a density function (now termed the
likelihood function) on x
k
. This density is then multiplied by the prior density P(x
k
)
and normalised to obtain the posterior distribution P(x
k
[ z
k
) describing knowledge in
the state after observation.
BAE SYSTEMS Proprietary Information, ASCTA use only 27
0
5
10
15
0
10
20
30
40
50
0
0.2
0.4
0.6
0.8
1
1.2
P(x
k
,x
k1
)dx
k
P(x
k1
)
x
k
x
k
=f(x
k1
,U
k
)
P(x
k1
,x
k
)
P(x
k
,x
k1
)dx
k1
x
k1
P(x
k
)
P(x
k
x
k1
)
Figure 7: Time update step for the full Bayes lter. At a time k1, knowledge of the state
x
k1
is summarised in a probability distribution P(x
k1
). A vehicle model, in the form
of a conditional probability density P(x
k
[ x
k1
), then describes the stochastic transition
of the vehicle from a state x
k1
at a time k 1 to a state x
k
at a time k. Functionally,
this state transition may be related to an underlying kinematic vehicle model in the
form x
k
= f(x
k1
, u
k
). The gure shows two typical conditional probability distributions
P(x
k
[ x
k1
) on the state x
k
given xed values of x
k1
. The product of this conditional
distribution with the marginal distribution P(x
k1
), describing the prior likelihood of
values of x
k
, gives the the joint distribution P(x
k
, x
k1
) shown as the surface in the
gure. The total marginal density P(x
k
) describes knowledge of x
k
after state transition
has occurred. The marginal density P(x
k
) is obtained by integrating (projecting) the joint
distribution P(x
k
, x
k1
) over all x
k1
. Equivalently, using the total probability theorem,
the marginal density can be obtained by integrating (summing) all conditional densities
P(x
k
[ x
k1
)weighted by the prior probability P(x
k1
) of each x
k1
. The process can
equally be run in reverse (a retroverse motion model) to obtain P(x
k1
) from P(x
k
) given
a model P(x
k1
[ x
k
).
BAE SYSTEMS Proprietary Information, ASCTA use only 28
where the last equality follows from the assumed independence of vehicle motion from map
and observations, and from the causality of the vehicle control input on vehicle motion.
Equation 37 is the time update step for the full Bayes data fusion algorithm. A
graphical description of this equation is shown in Figure 7.
The nal step is simply to substitute Equation 37 into Equation 36 to yield
P(x
k
[ Z
k
, U
k
, x
0
) = K.P(z
k
[ x
k
)
_
P(x
k
[ x
k1
, u
k
)P(x
k1
[ Z
k1
, U
k1
, x
0
)dx
k1
.
(38)
Equation 38 provides a recursive expression for the calculation of the joint posterior
P(x
j
[ Z
j
, U
j
, x
0
) for the state x
j
at a time j based on all observations Z
j
and all control
inputs U
j
up to and including time j. The recursion is a function of a state transition
model P(x
k
[ x
k1
, u
k
) and an observation model P(z
k
[ x
k
). A related problem is to in
clude the complete history of states (a smoothed trajectory estimates) as P(X
k
[ Z
k
, U
k
).
However this is not usually required.
2.2.5 Distributed Data Fusion with Bayes Theorem
...... ......
X
z
1
z
n
z
i
Fusion Centre
P(z
i
x)
Sensor i
P(z
1
x)
Sensor 1
P(z
n
x)
Sensor n
1
(x)
i
(x)
n
(x)
P(xZ
n
)=C P( x)
i
(x)
n
i=1
P(x)
Figure 8: The distributed implementation of the independent likelihood pool. Each sensor
maintains its own model in the form of a conditional probability distribution P
i
(z
i
[ x).
On arrival of a measurement, z
i
, the sensor model is instantiated with the associated
observation to form a likelihood
i
(x). This is transmitted to a central fusion centre were
the normalised product of likelihoods and prior yields the posterior distribution P(x [ Z
n
).
BAE SYSTEMS Proprietary Information, ASCTA use only 29
Providing the essential rules of conditional probabilities and Bayes Theorem are fol
lowed, it is not dicult to develop methods for distributing the data fusion problem.
Figure 8 shows one such distributed architecture. In this case, the sensor models, in the
form of likelihood functions, are maintained locally at each sensor site. When an obser
vation is made, these likelihoods are instantiated to provide a likelihood function
i
(x)
describing a probability distribution over the true state of the world. Importantly, it is
this likelihood that is transmitted to the fusion centre. Thus, the sensor talks to the cen
tre in terms of the underlying state of the world, and not in terms of the raw observation
(as is the case in Figure 4). This has the advantage that each sensor is modular and
talks in a common state language. However, it has the disadvantage that a complete
likelihood, rather than a single observation, must be communicated. The central fusion
center simply computes the normalised product of communicated likelihoods and prior to
yield a posterior distribution as
P(x [ Z
n
) = P(x)
i
(x)
Example 6
Continuing with example 3, the two sensors are now distributed in the form of Figure
8. The two sensor likelihood matrices P
1
(z
1
[ x);
z
1
z
2
z
3
x
1
0.45 0.45 0.1
x
2
0.45 0.45 0.1
x
3
0.1 0.45 0.8
and P
2
(z
2
[ x);
z
1
z
2
z
3
x
1
0.45 0.1 0.45
x
2
0.1 0.45 0.45
x
3
0.45 0.45 0.1
respectively, are stored locally with each sensor. As observations occur, these likelihood
functions are instantiated with observations to produce likelihood functions on x. For ex
ample, if z
1
= z
1
is observed with sensor 1, z
2
= z
1
is observed with sensor 2, the instan
tiated likelihood functions returned to the central processor are
1
(x) = (0.45, 0.45, 0.1)
and
2
(x) = (0.45, 0.1, 0.45) respectively. Note again, these are probability densities de
ned on the state. The communicated likelihoods are fused in the central processor by
simple multiplication with the prior. On successive observations, likelihoods continue to
be communicated by the local sensors.
A second approach to distributed data fusion using Bayes Theorem is shown in Figure
9. In this case, each sensor computes a likelihood but then combines this, locally, with
BAE SYSTEMS Proprietary Information, ASCTA use only 30
......
X
z
1
z
n
Fusion Centre
P(z
1
x)
Sensor 1
P(x
k
z
1
)
P
1
(x
k
z)
P(z
n
x)
Sensor n
P(x
k
z
n
)
P
n
(x
k
z)
P(x
k
Z
n
)
P(x
k
Z
n
)=C P( x
k1
)
n
i=1
P
i
(x
k
z)
P(x
k1
Z
n
)
k k1
Figure 9: A distributed implementation of the independent opinion pool in which each
sensor maintains both its own model and also computes a local posterior. The complete
posterior is made available to all sensors and so they become, in some sense, autonomous.
The gure shows Bayes Theorem in a recursive form.
BAE SYSTEMS Proprietary Information, ASCTA use only 31
the prior from the previous timestep, so producing a local posterior distribution on the
state. This is then communicated to the central fusion centre. The fusion centre then
recovers the new observation information by dividing each posterior by the communicated
(global) prior and then taking a normalised product to produce a new global posterior.
This posterior is then communicated back to the sensors and the cycle repeats in recursive
form. The advantage of this structure is that global information, in state form, is made
available at each local sensor site. This makes each sensor, in some sense, autonomous.
The disadvantage of this architecture is the need to communicate state distributions both
to and from the central fusion center. Distributed data fusion methods will be discussed
in detail later in this course.
Example 7
Consider the same two sensors as example 6 now distributed as Figure 9. Assuming an
existing prior, P(x) = (1/3, 1/3, 1/3) , communicated to the two sensors. After observa
tion of, say, z
1
= z
1
by sensor 1 and z
2
= z
1
by sensor 2, local likelihoods are instantiated
but then multiplied by the prior to yield local posteriors P
1
(x [ z
1
) = (0.45, 0.45, 0.1) and
P
2
(x [ z
2
) = (0.45, 0.1, 0.45) respectively. These posteriors are communicated to the cen
tral fusion site. At the fusion center, care must be taken to avoid double counting the
common prior information communicated in the posteriors from each sensor site. This is
achieved by dividing the posterior by the prior before fusion as
P
12
(x [ z
1
, z
2
) = P(x)
_
P
1
(x [ z
1
)
P(x)
_ _
P
2
(x [ z
2
)
P(x)
_
= (0.692, 0.154, 0.154)
where the division is taken element wise. In this form of distributed architecture, the
combined posterior is now communicated back to the local sensor sites. This information
can be used locally at the sensor site. If a second observation is made locally at the site,
the the local likelihood is combined with this global posterior to yield a new local posterior.
For example, if sensor 2 now detects a target type 1,then the local posterior is computed
from
P
12
(x [ z
1
, z
2
z
2
) = (0.108, 0.024, 0.868) = (0.692, 0.154, 0.154) (0.45, 0.1, 0.45).
This new posterior is then communicated to the central fusion site and the process repeats.
Almost any data fusion structure is possible as long as the essential rules of conditional
probability and Bayes theorem are followed. Figure 10 shows a Bayes lter in feedback
form. The lter outputs state predictions in the form of a pdf P(x
k
[ Z
k1
) of the state x
k
at time k given all observations Z
k1
up to time k 1 (typically using Equation 37). This
is essentially a prior distribution and provides a distribution on the state at time k prior
to an observation being made at this time (and thus is marked with a superscript to
denote before update). The prior is fedback and multiplied by the likelihood (x) as
sociated with an incoming observation. This multiplication essentially implements Bayes
BAE SYSTEMS Proprietary Information, ASCTA use only 32
z
P(x

k+1
Z
k
)
P( zx)
P( x

x
+
)
P(x
+
k
Z
k
)
P( x

k
Z
k1
) k k1
Figure 10: A Bayesian formulation of the classical feedback aided navigation system,
fusing rate data with external observation information (see text for details).
Theorem to produce the posterior P(x
+
k
[ Z
k
), where the superscript + denotes after
observation. The likelihood itself is generated from the sensor model P(z [ x) in the nor
mal manner. The posterior is fedback and multiplied by another distribution P(x
[ x
+
)
which essentially predicts the future state on the basis of current state. Multiplying this
distribution by the fedback posterior generates a new prior. The cycle, after delay, then
repeats. This structure is a Bayesian implementation of the classical feedback aided
navigation system. The state prediction distribution P(x
[ x
+
) has the same role as an
inertial system, using integrated rate information to generate predictions of state. The
feedback loop is corrected by reference to some external observation (the aiding sensor)
modeled by the likelihood function P(z [ x). Thus rate and external information are fused
in a predictorcorrector arrangement.
2.2.6 Data Fusion with LogLikelihoods
In data fusion problems and in Bayes networks, it is often easier to work with the log of a
probability, rather than the probability itself. These loglikelihoods are more convenient
computationally than probabilities as additions and subtractions rather than multiplica
tions and divisions are employed in fusing probability information. Further, loglikelihoods
are also more closely related to formal denitions of information.
The loglikelihood or conditional loglikelihood are dened as;
l(x)
= log P(x), l(x [ y)
= log P(x [ y). (39)
It is clear that the loglikelihood is always less than zero and is only equal to zero when
all probability mass is assigned to a single value of x; l(x) 0. The loglikelihood itself
can sometimes be a useful and ecient means of implementing probability calculations.
For example, taking logs of both sides of Equation 12 we can rewrite Bayes theorem in
terms of loglikelihood as
l(x [ z) = l(z [ x) +l(x) l(z). (40)
Example 8
BAE SYSTEMS Proprietary Information, ASCTA use only 33
Consider again the twosensor discrete target identication example (Example 3). The
loglikelihood matrix for the rst sensor (using natural logs) is
z
1
z
2
z
3
x
1
0.799 2.303 0.799
x
2
2.303 0.799 0.799
x
3
0.799 0.799 2.303
and for the second
z
1
z
2
z
3
x
1
0.799 2.303 0.799
x
2
2.303 0.799 0.799
x
3
0.799 0.799 2.303
The posterior likelihood (given a uniform prior) following observation of target 1 by sensor
1 and target 1 by sensor 2 is the sum of the rst columns of each of the likelihood matrices
l(x [ z
1
, z
1
) = l
1
(z
1
[ x) +l
2
(z
1
[ x) +C
= (0.7985, 0.7985, 2.3026) + (0.7985, 2.3026, 0.7985) +C
= (1.5970, 3.1011, 3.1011) +C
= (0.3680, 1.8721, 1.8721)
where the constant C = 1.229 is found through normalisation (which in this case requires
that the antilogs sum to one). The rst thing to note is that the computation is obviously
simpler than in the case of obtaining products of probability distributions, particularly as
the dimension of the state vectors increase. The second thing is that the normalising
constant is additive. Thus, normalisation need only occur if probabilities are required,
otherwise, the relative magnitude of the loglikelihoods are sucient to indicate relative
likelihoods (the smaller the loglikelihood, the nearer the probability is to one).
Example 9
A second interesting example of loglikelihoods is in the case where prior and sensor
likelihoods are Gaussian as in Example 4. Taking logs of Equation 27, gives
l(x [ Z
k
) =
1
2
(x
k
x)
2
2
k
=
1
2
(z
k
x)
2
2
1
2
(x
k1
x)
2
2
k1
+C. (41)
as before, completing squares gives
x
k
=
2
k1
2
k1
+
2
z
k
+
2
2
k1
+
2
x
k1
,
BAE SYSTEMS Proprietary Information, ASCTA use only 34
and
2
k
=
2
2
k1
2
+
2
k1
,
thus the loglikelihood is quadratic in x; for each value of x, a loglikelihood is specied as
1
2
(x
k
x)
2
2
k
, modulo addition of a constant C.
1
N
2
log P( x{Z
k1
})
log P( x{Z
k
})
log P(z
1
(k) x)
log P(z
N
(k) x)
log P(z
2
(k) x)
z
1
(k)
z
2
(k)
z
N
(k)
k k1
Sensor
Central Processor Sensor Models
Figure 11: A loglikelihood implementation of a fully centralised data fusion architecture.
1
N
2
log P( x{Z
k1
})
log P( x{Z
k
})
log P(z
1
(k) x)
log P(z
N
(k) x)
log P(z
2
(k) x)
z
1
(k)
z
2
(k)
z
N
(k)
k k1
Sensor
Central Processor
Sensor Models
Figure 12: A loglikelihood implementation of the independent likelihood pool architec
ture.
Loglikelihoods are a convenient way of implementing distributed data fusion architec
tures. Figure 11 shows a loglikelihood implementation of a fully centralised data fusion
architecture (equivalent to Figure 4) in which observations are transmitted directly to
a central fusion centre which maintains the sensor likelihood functions. Fusion of infor
mation is simply a matter of summing loglikelihoods. Figure 12 shows a loglikelihood
implementation of the independent likelihood architecture (equivalent to Figure 8). In
this case, each sensor maintains its own model and communicates loglikelihoods to a
BAE SYSTEMS Proprietary Information, ASCTA use only 35
Central Processor
log P(z
1
(k) x)
z
1
(k)
k k1
log P( x{Z
k
}
1
)
Sensor 1
log P(z
N
(k) x)
z
N
(k)
k k1
log P( x{Z
k
}
N
)
Sensor N
log P(z
2
(k) x)
z
2
(k)
k k1
log P( x{Z
k
}
2
)
Sensor 2
log P( x{Z
k
})
Figure 13: A loglikelihood implementation of the independent opinion pool architecture.
central processor. The fusion of loglikelihoods is again simply a summation. Figure 13
shows a loglikelihood implementation of the independent opinion pool (equivalent to Fig
ure 9) in which each local sensor maintains a posterior likelihood which is communicate
to the central fusion center. In all cases, the loglikelihood implementation involves only
simple addition and subtraction in the form of classical feedback loops. These can be
easily manipulated to yield a wide variety of dierent architectures.
2.3 Information Measures
Probabilities and loglikelihoods are dened on states or observations. It is often valuable
to also measure the amount of information contained in a given probability distribution.
Formally, information is a measure of the compactness of a distribution; logically if a
probability distribution is spread evenly across many states, then its information content
is low, and conversely, if a probability distribution is highly peaked on a few states, then
its information content is high. Information is thus a function of the distribution, rather
than the underlying state. Information measures play an important role in designing
and managing data fusion systems. Two probabilistic measures of information are of
particular value in data fusion problems; the Shannon information (or entropy) and the
Fisher information.
2.3.1 Entropic Information
The entropy or Shannon information
4
H
P
(x) associated with a probability distribution
P(x), dened on a random variable x, is dened as the expected value of minus the
4
Properly, information should be dened as the negative of entropy; when entropy is a minimum,
information is a maximum. As is usual, we shall ignore this distinction and usually talk about entropy
minimization when we really mean information maximisation.
BAE SYSTEMS Proprietary Information, ASCTA use only 36
loglikelihood. For continuousvalued random variables this is given by (see [46] Chapter
15)
H
P
(x)
= Elog P(x) =
_
xA
P(x) log P(x). (43)
Note that following convention, we have used x as an argument for H
P
() even though
the integral or sum is taken over values of x so H
P
() is not strictly a function of x but
is rather a function of the distribution P().
The entropy H
P
() measures the compactness of a density P(). It achieves a minimum
of zero when all probability mass is assigned to a single value of x; this agrees with an
intuitive notion of a most informative distribution. Conversely, when probability mass
is uniformly distributed over states, the entropy in the distribution is a maximum; this
too agrees with the idea of least informative (maximum entropy) distributions. Maximum
entropy distributions are often used as prior distributions when no useful prior information
is available. For example, if the random variable x can take on at most n discrete values
in the set A, then the least informative (maximum entropy) distribution on x is one
which assigns a uniform probability 1/n to each value. This distribution will clearly have
an entropy of log n. When x is continuousvalued, the least informative distribution is
also uniform. However, if the range of x is continuous then the distribution is technically
not well dened as probability mass must be assigned equally over an innite range. If
the information is to be used as a prior in Bayes rule this technical issue is often not a
problem as P(x) can be set to any convenient constant value over the whole range of
x;
5
P(x) = 1, x, for example, without aecting the values computed for the posterior
P(x [ z).
It can be shown that, up to a constant factor and under quite general conditions of
preference ordering and preference boundedness, this denition of entropy is the only rea
sonable denition of informativenss. An excellent proof of this quite remarkable result
(rst shown by Shannon) can be found in [15]. The implications of this in data fusion
problems are manyfold. In particular it argues that entropy is a uniquely appropri
ate measure for evaluating and modeling information sources described by probabilistic
models. Such ideas will be examined in later sections on system performance measures,
organization and management.
2.3.2 Conditional Entropy
The basic idea of entropy can logically be extended to include conditional entropy; for
continuousvalued random variables
H
P
(x [ z
j
)
= Elog P(x [ z
j
) =
_
P(x [ z
j
) log P(x [ z
j
)dx (44)
5
A distribution such as this which clearly violates the constraint that
_
+
P(x)dx = 1 is termed an
improper distribution, or in this case an improper prior.
BAE SYSTEMS Proprietary Information, ASCTA use only 37
and for discrete random variables
H
P
(x [ z
j
)
= Elog P(x [ z
j
) =
x
P(x [ z
j
) log P(x [ z
j
). (45)
This should be interpreted as the information (entropy) about the state x contained in
the distribution P( [ z) following an observation z
j
. Note that H
P
(x [ z) is still a function
of z, and thus depends on the observation made.
The mean conditional entropy, H(x [ z), taken over all possible values of z, is given
by
H(x [ z)
= EH(x [ z)
=
_
+
P(z)H(x [ z)dz
=
_
+
_
+
_
+
z
H(x [ z)P(z)
=
x
P(x, z) log P(x [ z). (47)
for discrete random variables. Note that H(x [ z) is not a function of either x or z. It
is a essentially a measure of the information that will be obtained (on the average) by
making an observation before the value of the observation is known.
Example 10
Recall example 3 of two sensors observing a discrete state; type 1 target, type 2 target,
and no target. Consider rst sensor 1. Assuming a uniform prior distribution, the infor
mation (entropy) obtained about the state given that an observation z = z
1
has been made
(rst column of the likelihood matrix for sensor 1), is simply given by (using natural logs)
H
P
1
(x [ z
1
) =
3
i=1
P
1
(x
i
[ z
1
) log P
1
(x
i
[ z
1
) = 0.9489. (48)
For the rst sensor, the conditional entropy for each possible observation is
H
P
1
(x [ z) =
3
i=1
P
1
(x
i
[ z) log P
1
(x
i
[ z) = (0.9489, 0.9489, 0.6390). (49)
Thus, observing either z
1
or z
2
is equally as informative (indeed it provides exactly the
same information as sensor 1 can not distinguish between targets), but observing z
3
(no
BAE SYSTEMS Proprietary Information, ASCTA use only 38
target) is most informative because probability mass is relatively concentrated on the no
target state. Successive observation of the z
1
or z
2
with sensor 1, yields a posterior den
sity on x of (0.5, 0.5, 0.0) which has an information value (entropy) of log(2) = 0.6931.
Successive observation of z
3
yields a posterior of (0.0, 0.0, 1.0) which has an entropy of
zero (log(1), the minimum entropy).
Similar calculations for the likelihood matrix of sensor 2 give the conditional entropy
for each observation as H
P
2
(x [ z) = (0.948, 0.948, 0.948); that is any observation is equally
as informative. This is because, in the likelihood function for sensor 2, the relative distri
bution of probability mass is the same for any observation, even though the mass itself is
placed on dierent states. Successive observation of any of z
1
, z
2
, or z
3
results in probabil
ity being assigned to only two of three states and thus for the posterior entropy to achieve
a minimum of log(2) = 0.6931. Note that, as with sensor 1 observing z
1
or z
2
, entropy
achieves a lower bound not equal to zero.
To nd the mean conditional entropy, the joint distribution P(x, y) is required. This
is computed from
P(x, z) = P(z [ x)P(x)
=
_
_
0.45 0.45 0.1
0.45 0.45 0.1
0.1 0.1 0.8
_
_
_
_
1/3
1/3
1/3
_
_
=
_
_
0.1500 0.1500 0.0333
0.1500 0.1500 0.0333
0.0333 0.0333 0.2667
_
_ ,
(Note the sum of elements is equal to 1). Substituting these values into Equation 47 and
summing over both x and z gives the mean conditional entropy as 0.8456.
Example 11
It is of considerable future interest to provide a measure for the entropy of a Gaussian
(normal) distribution. The pdf for an ndimensional Gaussian is given by:
P(x) = N(x, P) = [ 2P [
1/2
exp
_
1
2
(x x)
T
P
1
(x x)
_
, (50)
where x is the mean of the distribution and P the covariance, and where [ [ denotes
matrix determinant. The entropy for this distribution is obtained as follows
H
P
(x) = Elog P(x)
=
1
2
E(x x)
T
P
1
(x x) + log[(2)
n
[ P []
=
1
2
E
ij
(x
i
x
i
)P
1
ij
(x
j
x
j
)
1
2
log[(2)
n
[ P []
=
1
2
ij
E(x
j
x
j
)(x
i
x
i
) P
1
ij
1
2
log[(2)
n
[ P []
BAE SYSTEMS Proprietary Information, ASCTA use only 39
=
1
2
i
P
ji
P
1
ij
1
2
log[(2)
n
[ P []
=
1
2
j
(PP
1
)
jj
1
2
log[(2)
n
[ P []
=
1
2
j
1
jj
1
2
log[(2)
n
[ P []
=
n
2
1
2
log[(2)
n
[ P []
=
1
2
log[(2e)
n
[ P []. (51)
Thus the entropy of a Gaussian distribution is dened only by the state vector length n
and the covariance P. The entropy is proportional to the log of the determinant of the
covariance. The determinant of a matrix is a volume measure (recall that the determinant
is the product of the eigenvalues of a matrix and the eigenvalues dene axis lengths in n
space). Consequently, the entropy is a measure of the volume enclosed by the covariance
matrix and consequently the compactness of the probability distribution.
If the Gaussian is scalar with variance
2
, then the entropy is simply given by
H(x) = log
2e
For a two random variables x and z which are jointly Gaussian, with correlation
zx
=
xz
2
xx
2
zz
, the conditional entropy is given by
H(x [ z) = log
xx
_
2e(1
2
xz
).
Thus when the variables are uncorrelated,
xz
= 0, the conditional entropy is just the
entropy in P(x), H(x [ z) = H(x), as z provides no additional information about x.
Conversely, when the variables are highly correlated
xz
1, the conditional entropy goes
to zero as complete knowledge of z implies complete knowledge of x.
2.3.3 Mutual Information
With these denitions of entropy and conditional entropy, it is possible to write an infor
mation form of Bayes theorem. Taking expectations of Equation 40 with respect to both
the state x and the observation z gives (we will now drop the sux P when the context
of the distribution is obvious)
H(x [ z) = H(z [ x) +H(x) H(z). (52)
Simply, this describes the change in entropy or information following an observation from
a sensor modeled by the likelihood P(z [ x).
BAE SYSTEMS Proprietary Information, ASCTA use only 40
Being able to describe changes in entropy leads naturally to asking an important
question; what is the most informative observation I can make? This question may be
answered through the idea of mutual information.
The mutual information I(x, z) obtained about a random variable x with respect to
a second random variable z is now dened as
I(x, z) = Elog
P(x, z)
P(x)P(z)
= Elog
P(x [ z)
P(x)
= Elog
P(z [ x)
P(z)
(53)
Mutual information is an a priori measure of the information to be gained through ob
servation. It is a function of the ratio of the density P(x [ z) following an observation to
the prior density P(x). The expectation is taken over z and x, so the mutual information
gives an average measure of the gain to be expected before making the observation. If the
underlying probability distributions are continuous then
I(x, z) =
_
+
_
+
P(x, z) log
P(x, z)
P(x)P(z)
dxdz, (54)
and for discrete distributions
I(x, z) =
zZ
xA
P(x, z) log
P(x, z)
P(x)P(z)
. (55)
Noting that
P(x, z)
P(x)P(z)
=
P(x [ z)
P(x)
, (56)
then if x and z are independent, the expressions in Equation 56 become equal to one and
(taking logs) the mutual information becomes equal to zero. This is logical; if knowledge
of the state is independent of the observation, the information to be gained by taking an
observation (the mutual information) is zero. Conversely, as x becomes more dependent on
z, then P(x [ z) becomes more peaked or compact relative to the prior distribution P(x)
and so mutual information increases. Note that mutual information is always positive (it
is not possible to lose information by taking observations).
Equation 53 can be written in terms of the component entropies as
I(x, z) = H(x) +H(z) H(x, z)
= H(x) H(x [ z)
= H(z) H(z [ x) (57)
Equation 57 measures the compression of the probability mass caused by an observation.
Mutual information provides an average measure of how much more information we would
BAE SYSTEMS Proprietary Information, ASCTA use only 41
have about the random variable x if the value of z where known. Most importantly mutual
information provides a preexperimental measure of the usefulness of obtaining information
(through observation) about the value of z.
Example 12
Continuing the two sensors target identication example (Example 10), it is interesting
to see what value mutual information has in deciding which sensor should be used for
correctly identifying a target (this is a sensor management problem).
First assume that the prior information on x is uniform so that
P(x) = [1/3, 1/3, 1/3]
T
.
Consider taking an observation with sensor 1. The total probability of observation P(z)
can be found by summing the joint probability P(x, z) (computed in Example 10) to obtain
P(z) = [1/3, 1/3, 1/3]. The mutual information, I(x, z), on using sensor 1 to take an
observation is then
I(x, z) =
j
P(z
j
) log P(z
j
)
_
i
P(x
i
, z
j
) log P(z
j
[ x
i
)
_
= 1.0986 0.8456 = 0.2530.
Suppose we now go ahead and take an observation with sensor 1 and the result is z
1
, an
observation of target 1. The prior is now updated as (Example 2)
P(x) = P
1
(x [ z
1
) = [0.45, 0.45, 0.1].
The mutual information gain from using sensor 1 given this new prior is I(x, z) = 0.1133.
Note, that this is smaller than the mutual information obtained with a uniform prior.
This is because we now have observation information and so the value of obtaining more
information is less. Indeed, mutual information shows that the more we observe, the less
value there is in taking a new observation. In the limit, if we repeatedly observe z
1
with
sensor 1, then the prior becomes equal to [0.5, 0.5, 0.0] (Example 2). In this case the
predicted mutual information gain is zero. That is, there is no value in using sensor 1,
yet again, to obtain information.
Similarly, again assuming a uniform prior, the mutual information gain predicted
for using sensor 2 is 0.1497.This is lower than sensor 1, when the notarget density is
more highly peaked. If we observe target 1 with sensor 2 then the updated prior becomes
P(x) = [0.45, 0.1, 0.45]. Using this prior, the mutual information gain for using sensor
2 is only 0.1352; smaller, but not as small as the corresponding mutual information gain
from sensor 1 after the rst measurement. Continual observation of target 1 with sensor
2 results in the posterior becoming equal to [0.5, 0.0, 0.5] and if, this is substituted into
the equation for mutual information gain, we obtain a predicted mutual information gain
of 0.1205; even after taking an innite number of measurements. The reason for this
is subtle; as sensor 2 can not distinguish targets, any observation other than that of
target 1 will still provide additional information and as mutual information is an average
measure, it is still predicts, on the average, an increase in information (even though further
BAE SYSTEMS Proprietary Information, ASCTA use only 42
observation of target 1 will not provide any more information). Continual observation of
target 1 with sensor 2 would never happen in practice. As the likelihood suggests, if the
true state were target 1, then sensor 2 would be as likely to not to detect a target at all
as report a target 1. Indeed, if we mix detections as the likelihood suggests, the posterior
tends to [1, 0, 0] or [0, 1, 0] which then suggests zero mutual information gain from sensor
2.
Now, suppose sensor 1 is used to obtain the rst measurement and target 1 is detected
so that the new prior is [0.45, 0.45, 0.1]. Which of sensor 1 or 2 should be used to make
the second observation ? With this prior, the predicted mutual information gain from
using sensor 1 is 0.1133, and from using sensor 2 is 0.1352. This (logically) tells us
to use sensor 2. Now sensor 2 is equally as likely to detect the correct target or return
a nodetect. Sensor 2 is now employed and a nodetect is returned yielding a posterior
[0.4880, 0.4880, 0.0241]. With this as a prior, mutual information will tell us to continue
with sensor 2 until we get a return for either target 1 or 2. If this (target 2, say) happens
next time round, then the posterior will be [0.1748, 0.7864, 0.0388]; we now have, for the
rst time a target preference (2) and sensor 1 and 2 can both be used to rene this estimate.
With this prior, the mutual information for sensor 1 is 0.0491 and for sensor 2 is 0.0851;
thus sensor 2 will be used on the fourth sensing action.
This process of predicting information gain, making a decision on sensing action, then
sensing, is an ecient and eective way of managing sensing resources and of determining
optimal sensing policies. However, it should be noted that this is an average policy. As
we saw with the behaviour of sensor 2, it is sometimes not the most logical policy if we
value things in a manner other than on the average. This issue will be dealt with in more
detail in the section on decision making.
Example 13
A second interesting application of mutual information gain is Example 5. In this
example, general probability distributions, dening target location, are dened on a location
grid. Entropy measures can be obtained for distributions on this grid from
H(x, y) = Elog P(x, y)
=
_ _
P(x, y) log P(x, y)dxdy
P(x
i
, y
j
) log P(x
i
, y
j
)x
i
y
j
(58)
The total entropy associated with the prior shown in Figure 5(a), computed with Equation
58 is H(x) = 12.80. The total entropy associated with the sensor 1 likelihood shown in
Figure 5(b) is H(z [ x) = 4.2578, and the entropy associated with the resulting posterior
of Figure 5(c) is H(x [ z) = 3.8961. The predicted mutual information gain from this
observation is therefore I(x, z) = 8.9039 (a big gain). However, the predicted mutual
information gain from a second observation using sensor 1 is only 0.5225. This is reected
in the slight change in posterior from Figure 5(c) to Figure 5(d). Conversely, the predicted
BAE SYSTEMS Proprietary Information, ASCTA use only 43
mutual information gain from sensor 2, with likelihood shown in Figure 5(e), is 2.8598 (a
relatively large gain). This is reected in the posterior shown in Figure 5(f ) which has an
entropy of only 0.5138.
This example shows that it is straightforward to apply principles of entropy to any
probability type of probabilistic information; to measure compactness of information and
the prediction of information gain using mutual information.
2.3.4 Fisher Information
A second measure of information commonly used in probabilistic modeling and estimation
is the Fisher information. Unlike Shannon information, Fisher information may only be
dened on continuous distributions. The Fisher information J(x) is dened as the second
derivative of the loglikelihood
6
J(x) =
d
2
dx
2
log P(x). (59)
In general, if x is a vector, then J(x) will be a matrix, usually called the Fisher Information
Matrix. The Fisher information describes the information content about the values of
x contained in the distribution P(x). The Fisher information measures the surface of a
bounding region containing probability mass. Thus, like entropy, it measures compactness
of a density function. However, entropy measures a volume and is thus a single number,
whereas Fisher information is a series of numbers (and generally a matrix) measuring the
axes of the bounding surface.
The Fisher information is particularly useful in the estimation of continuous valued
quantities. If P(x) describes all the (probabilistic) information we have about the quantity
x, then the smallest variance that we can have on an estimate of the true value of x is
known as the CramerRao lower bound, and is equal to the Fisher information J(x). An
estimator that achieves this lower bound is termed ecient.
Example 14
The simplest example of Fisher information is the information associated with a vector
x known to be Gaussian distributed with mean x and covariance P;
P(x) = N(x; x; P) =
1
(2)
n/2
[P[
exp
_
1
2
(x x)P
1
(x x)
T
_
(60)
Taking logs of this distribution, and dierentiating twice with respect to x gives J(x) =
P
1
. That is, the Fisher information is simply the inverse covariance. This agrees with
intuition; Gaussian distributions are often drawn as a series of ellipsoids containing prob
ability mass. P
1
describes the surface of this ellipsoid, the square root of its eigenvalues
being the dimensions of each axis of the ellipsoid.
6
The rst derivative of the loglikelihood is called the score function
BAE SYSTEMS Proprietary Information, ASCTA use only 44
The Fisher information plays an important role in multisensor estimation problems.
In conventional estimation problems when only a single source of information is used to
obtain information about an unknown state, it is common to talk of an estimate of this
state together with some associated uncertainty (normally a variance). However, in multi
sensor estimation problems it is dicult to describe any (statistical) relations there may
be between the dierent estimates produced by dierent combinations of sensors. This
problem can only be overcome by dealing directly with the likelihood functions associ
ated with the observations themselves and by explicitly accounting for any dependencies
between dierent estimates. The Fisher information provides a direct means of account
ing for these dependencies as it makes explicit the information available in the likelihood
function. We will return to this again in the Chapter on multisensor estimation.
2.3.5 The Relation between Shannon and Fisher Measures
The question arises that if entropy is considered to be the only reasonable measure of
information content in a distribution, why consider Fisher information at all as it must,
by denition, be an unreasonable measure of information. Fortunately, there is a sensi
ble explanation for this problem, in that for continuous variables, Fisher information and
Shannon information are indeed related by the loglikelihood function. Broadly, entropy
is related to the volume of a set (formally a typical set) containing a specied proba
bility mass. Fisher information is related in the surface area of this typical set. Thus
maximization of Fisher information is equivalent to minimization of entropy. A detailed
development of this relation using the Asymptotic Equipartion Property (AEP) is given
in Cover [17].
Example 15
If the pdf associated with a random variable x is known to be Gaussian distributed
with mean x and covariance P (as Equation 60), then an explicit relation for the relation
between Fisher and Shannon information can be obtained, and indeed is given in Equation
51 (Example 11) as;
H(x) =
1
2
log[(2e)
n
[ P []
This clearly shows the relation between the Fisher surface measure of information P
1
and the entropic volume measure through the determinant of P.
BAE SYSTEMS Proprietary Information, ASCTA use only 45
3 MultiSensor Estimation
Estimation is the single most important problem in sensor data fusion. Fundamentally,
an estimator is a decision rule which takes as an argument a sequence of observations and
whose action is to compute a value for the parameter or state of interest. Almost all data
fusion problems involve this estimation process: we obtain a number of observations from
a group of sensors and using this information we wish to nd some estimate of the true
state of the environment we are observing. Estimation encompasses all important aspects
of the data fusion problem. Sensor models are required to understand what information is
provided, environment models are required to relate observations made to the parameters
and states to be estimated, and some concept of information value is needed to judge
the performance of the estimator. Dening and solving an estimation problem is almost
always the key to a successful data fusion system.
This section begins with a brief summary of the Kalman lter algorithm. The intention
is to introduce notation and key data fusion concepts; prior familiarity with the basic
Kalman Filter algorithm is assumed (see either [20] or the numerous excellent books on
Kalman ltering [12, 5, 37]). The multisensor Kalman lter is then discussed. Three main
algorithms are considered; the groupsensor method, the sequential sensor method and
the inverse covariance form. The tracktotrack fusion algorithm is also described. The
problem of multipletarget tracking and data association are described. The three most
important algorithms for data association are introduced. Finally, alternative estimation
methods are discussed. In particular, maximum likelihood lters and various probability
distribution oriented methods. Subsequent sections consider the distributed Kalman lter
and dierent data fusion architectures.
3.1 The Kalman Filter
The Kalman Filter is a recursive linear estimator which successively calculates an estimate
for a continuous valued state, that evolves over time, on the basis of periodic observations
that of this state. The Kalman Filter employs an explicit statistical model of how the
parameter of interest x(t) evolves over time and an explicit statistical model of how the
observations z(t) that are made are related to this parameter. The gains employed in a
Kalman Filter are chosen to ensure that, with certain assumptions about the observation
and process models used, the resulting estimate x(t) minimises meansquared error
L(t) =
_
(x(t) x(t))
T
(x(t) x(t)) P(x(t) [ Z
t
)dx. (61)
Dierentiation of Equation 61 with respect to x(t) and setting equal to zero gives
x(t) =
_
x(t)P(x(t) [ Z
t
)dx, (62)
which is simply the conditional mean x(t) = Ex(t) [ Z
t
. The Kalman lter, and indeed
any meansquarederror estimator, computes an estimate which is the conditional mean;
an average, rather than a most likely value.
BAE SYSTEMS Proprietary Information, ASCTA use only 46
The Kalman lter has a number of features which make it ideally suited to dealing
with complex multisensor estimation and data fusion problems. In particular, the explicit
description of process and observations allows a wide variety of dierent sensor models
to be incorporated within the basic algorithm. In addition, the consistent use statistical
measures of uncertainty makes it possible to quantitatively evaluate the role each sensor
places in overall system performance. Further, the linear recursive nature of the algorithm
ensure that its application is simple and ecient. For these reasons, the Kalman lter
has found widespread application in many dierent data fusion problems [53, 3, 5, 37].
3.1.1 State and Sensor Models
The starting point for the Kalman lter algorithm is to dene a model for the states to
be estimated in the standard statespace form;
x(t) = F(t)x(t) +B(t)u(t) +G(t)v(t), (63)
where
x(t) 1
n
is the state vector of interest,
u(t) 1
s
is a known control input,
v(t) 1
q
is a random variable describing uncertainty in the evolution of the state,
F(t) is the n n state (model) matrix,
B(t) is the n s input matrix, and
G(t) is the n q noise matrix.
An observation (output) model is also dened in standard statespace form;
z(t) = H(t)x(t) +D(t)w(t), (64)
where
z(t) 1
m
is the observation vector,
w(t) 1
r
is a random variable describing uncertainty in the observation,
H(t) is the mn observation (model) matrix,
D(t) is the mr observation noise matrix.
These equations dene the evolution of a continuoustime system with continuous obser
vations being made of the state. However, the Kalman lter is almost always implemented
in discretetime. It is straightforward to obtain a discretetime version of Equations 63
and 64.
BAE SYSTEMS Proprietary Information, ASCTA use only 47
First, a discretetime set t = t
0
, t
1
, t
k
, is dened. Equation 64 can be written
in discrete time as
z(t
k
) = H(t
k
)x(t
k
) +D(t
k
)w(t
k
), t
k
t (65)
where z(t
k
), x(t
k
) and w(t
k
) are the discretetime observation, state and noise vectors
respectively, and H(t
k
) and D(t
k
) the observation and noise models evaluated at the
discrete time instant t
k
. The discretetime form of the state equation requires integration
of Equation 63 over the interval (t
k
, t
k1
) as
x(t
k
) = (t
k
, t
k1
)x(t
k1
) +
_
t
k
t
k1
(t
k
, )B()u()d +
_
t
k
t
k1
(t
k
, )G()v()d. (66)
where (, ) is the state transition matrix satisfying the matrix dierential equation
(t
k
, t
k1
) = F(t
k
)(t
k
, t
k1
), (t
k1
, t
k1
) = 1. (67)
The state transition matrix has three important properties that should be noted:
1. It is uniquely dened for all t, t
0
in [0, ].
2. (The semigroup property) (t
3
, t
1
) = (t
3
, t
2
)(t
2
, t
1
).
3. (t
k
, t
k1
) is non singular and
1
(t
k
, t
k1
)=(t
k1
, t
k
).
When F(t) = F is a constant matrix, the state transition matrix is given by
(t
k
, t
k1
) = (t
k
t
k1
) = exp F(t
k
t
k1
). (68)
which clearly satises these three properties.
If u(t) = u(t
k
) and v(t) = v(t
k
) remain approximately constant over the interval
(t
k1
, t
k
) then the following discretetime models can be dened;
F(t
k
)
= (t
k
, t
k1
)
B(t
k
)
=
_
t
k
t
k1
(t
k
, )B()d
G(t
k
)
=
_
t
k
t
k1
(t
k
, )G()d.
(69)
Equation 66 can then be written in a discretetime from equivalent to Equation 63 as
x(t
k
) = F(t
k
)x(t
k1
) +B(t
k
)u(t
k
) +G(t
k
)v(t
k
). (70)
The accuracy of this model could be improved by taking mean values for both u(t) and
v(t) over the sampling interval.
BAE SYSTEMS Proprietary Information, ASCTA use only 48
In almost all cases the time interval t(k)
= t
k
t
k1
between successive samples of
the state remains constant. In this case it is common particle to drop the time argument
and simply index variables by the sample number. In this case Equation 70 is written as
x(k) = F(k)x(k 1) +B(k)u(k) +G(k)v(k), (71)
and Equation 65 as
z(k) = H(k)x(k) +D(k)w(k) (72)
Equations 71 and 72 are the model forms that will be used throughout this section unless
discussion of asynchronous data is relevant.
A basic assumption in the derivation of the Kalman lter is that the random sequences
v(k) and w(k) describing process and observation noise are all Gaussian, temporally
uncorrelated and zeromean
Ev(k) = Ew(k) = 0, k, (73)
with known covariance
Ev(i)v
T
(j) =
ij
Q(i), Ew(i)w
T
(j) =
ij
R(i). (74)
It is also generally assumed that the process and observation noises are also uncorrelated
Ev(i)w
T
(j) = 0, i, j. (75)
These are eectively equivalent to a Markov property requiring observations and successive
states to be conditionally independent. If the sequences v(k) and w(k) are temporally
correlated, a shaping lter can be used to whiten the observations; again making the
assumptions required for the Kalman lter valid [37]. If the process and observation noise
sequences are correlated, then this correlation can also be accounted for in the Kalman
lter algorithm [1]. If the sequence is not Gaussian, but is symmetric with nite moments,
then the Kalman lter will still produce good estimates. If however, the sequence has a
distribution which is skewed or otherwise pathological, results produced by the Kalman
lter will be misleading and there will be a good case for using a more sophisticated
Bayesian lter [55]. Problems in which the process and observation models are nonlinear
are dealt in Section 3.1.5
We will use the following standard example of constantvelocity particle motion as the
basis for many of the subsequent examples on tracking and data fusion:
Example 16
Consider the linear continuoustime model for the motion of particle moving with
approximately constant velocity:
_
x(t)
x(t)
_
=
_
0 1
0 0
_ _
x(t)
x(t)
_
+
_
0
v(t)
_
.
BAE SYSTEMS Proprietary Information, ASCTA use only 49
0 10 20 30 40 50 60 70 80 90 100
5
0
5
10
15
20
25
Time(s)
P
o
s
i
t
i
o
n
(
m
)
True Target Track and Observations of Position
Target True Position
Target Observations
Figure 14: Plot of true target position and observations made of target position for the
constant velocity target model described in Example 16, with
q
= 0.01m/s and
r
= 0.1m
In this case the statetransition matrix from Equation 68 over a time interval T is given
by
(T) =
_
1 T
0 1
_
.
On the assumption that the process noise is white and uncorrelated with E[v(t)] = 0 and
E[v(t)v()] =
2
q
(t)(t ), then the equivalent discretetime noise process is given by
v(k) =
_
T
0
_
1
_
v(kT +)d =
_
T
2
/2
T
_
v(k).
With the denitions given in Equation 69, the equivalent discretetime model is given by
_
x(k)
x(k)
_
=
_
1 T
0 1
_ _
x(k 1)
x(k 1)
_
+
_
T
2
/2
T
_
v(k).
If observations are made at each time step k of the location of the particle the observation
model will be in the form
z
x
= [ 1 0 ]
_
x(k)
x(k)
_
+w(k), Ew
2
(k) =
2
r
BAE SYSTEMS Proprietary Information, ASCTA use only 50
Figure 14 shows a typical constantvelocity target motion generated according to these
models. The target position executes a random walk. Observations are randomly dispersed
around true target position.
These equations can trivially be extended to two (and three) dimensions giving a two
dimensional model in the form:
_
_
x(k)
x(k)
y(k)
y(k)
_
_
=
_
_
1 T 0 0
0 1 0 0
0 0 1 T
0 0 0 1
_
_
_
_
x(k 1)
x(k 1)
y(k 1)
y(k 1)
_
_
+
_
_
T
2
/2 0
T 0
0 T
2
/2
0 T
_
_
_
v
x
(k)
v
y
(k)
_
, (76)
and
_
z
x
z
y
_
=
_
1 0 0 0
0 0 1 0
_
_
_
x(k)
x(k)
y(k)
y(k)
_
_
+
_
w
x
(k)
w
y
(k)
_
(77)
with
R(k) = Ew(k)w
T
(k) =
_
2
rx
0
0
2
ry
_
.
This motion model is widely used in target tracking problems (in practice as well as the
ory) as it is simple, linear and admits easy solution for multipletarget problems. Figure
15 shows a typical x y target plot generated using this model. Target velocity and head
ing can be deduced from the magnitude and orientation of the estimated velocity vector
[ x(k), y(k)]
T
(Figure 15 (c) and (d)). The plots show that this simple model is capable of
accommodating many reasonable motion and maneuver models likely to be encountered in
typical target tracking problems. It should however be noted that this track is equivalent to
running independent models for both x and y as shown in Figure 14. In reality, motions
in x and y directions will be physically coupled, and therefore correlated, by the heading
of the target. This (potentially valuable information) is lost in this target model.
As this example shows, it is always best to start with a continuoustime model for the
state and then construct a discrete model, rather than stating the discrete model at the
outset. This allows for correct identication of noise transfers and noise correlations.
3.1.2 The Kalman Filter Algorithm
The Kalman lter algorithm produces estimates that minimise meansquared estimation
error conditioned on a given observation sequence
x(i [ j) = arg min
x(i [ j)
n
E(x(i) x)(x(i) x)
T
[ z(1), , z(j) . (78)
As has been previously demonstrated (Equation 62) the estimate obtained is simply the
expected value of the state at time i conditioned on the observations up to time j. The
BAE SYSTEMS Proprietary Information, ASCTA use only 51
0 1 2 3 4 5 6 7 8 9 10
x 10
4
1
1.05
1.1
1.15
1.2
1.25
1.3
1.35
1.4
1.45
x 10
4
Xposition (m)
Y
P
o
s
it
io
n
(
m
)
True Target Track
(a)
2.4 2.6 2.8 3 3.2 3.4 3.6 3.8 4
x 10
4
1.14
1.15
1.16
1.17
1.18
1.19
1.2
1.21
1.22
1.23
x 10
4
Xposition (m)
Y
P
o
s
it
io
n
(
m
)
True Target Track and Observations of Position
Target True Position
Target Observations
(b)
0 50 100 150 200 250 300 350 400 450 500
150
160
170
180
190
200
210
220
Time(s)
V
e
lo
c
it
y
(
m
/
s
)
True Target Velocity
(c)
0 50 100 150 200 250 300 350 400 450 500
0.25
0.2
0.15
0.1
0.05
0
0.05
0.1
0.15
0.2
0.25
Time(s)
H
e
a
d
in
g
(
r
a
d
s
)
True Target Heading
(d)
Figure 15: True target track for linear the uncoupled linear target model of Example 16
(a) x y track; (b) Detail of track and corresponding observations; (c) Deduced target
velocity; (d) Deduced target heading.
BAE SYSTEMS Proprietary Information, ASCTA use only 52
estimate is thus dened as the conditional mean
x(i [ j)
= Ex(i) [ z(1), , z(j)
= Ex(i) [ Z
j
. (79)
The estimate variance is dened as the mean squared error in this estimate
P(i [ j)
= E(x(i) x(i [ j))(x(i) x(i [ j))
T
[ Z
j
. (80)
The estimate of the state at a time k given all information up to time k will be written
as x(k [ k). The estimate of the state at a time k given only information up to time k 1
is called a onestepahead prediction (or just a prediction) and is written as x(k [ k 1).
The Kalman lter algorithm is now stated without proof. Detailed derivations can
be found in many books on the subject, [37, 5] for example (see also [20]). The state is
assumed to evolve in time according to Equation 71. Observations of this state are made
at regular time intervals according to Equation 72. The assumptions about the noise
processes entering the system, as described by Equations 73, 74 and 75, are assumed true.
It is also assumed that an estimate x(k 1 [ k 1) of the state x(k 1) at time k 1
based on all observations made up to and including time k 1 is available, and that this
estimate is equal to the conditional mean of the true state x(k 1) conditioned on these
observations. The conditional variance P(k 1 [ k 1) in this estimate is also assumed
known. The Kalman lter then proceeds recursively in two stages:
Prediction: A prediction x(k [ k 1) of the state at time k and its covariance P(k [ k 1)
is computed according to
x(k [ k 1) = F(k) x(k 1 [ k 1) +B(k)u(k) (81)
P(k [ k 1) = F(k)P(k 1 [ k 1)F
T
(k) +G(k)Q(k)G
T
(k). (82)
Update: At time k an observation z(k) is made and the updated estimate x(k [ k) of
the state x(k), together with the updated estimate covariance P(k [ k) is computed from
the state prediction and observation according to
x(k [ k) = x(k [ k 1) +W(k) (z(k) H(k) x(k [ k 1)) (83)
P(k [ k) = (1 W(k)H(k))P(k [ k 1)(1 W(k)H(k))
T
+W(k)R(k)W
T
(k) (84)
where the gain matrix W(k) is given by
W(k) = P(k [ k 1)H(k)
_
H(k)P(k [ k 1)H
T
(k) +R(k)
_
1
(85)
The Kalman lter is recursive or cyclic (see Figure 16). We start with an estimate,
generate a prediction, make an observation, then update the prediction to an estimate.
The lter makes explicit use of the process model in generating a prediction and the
BAE SYSTEMS Proprietary Information, ASCTA use only 53
State Prediction
x(kk1)=
F(k)x(k1k1)+G(k)u(k)
State Estimate at t
k1
x(k1k1)
Measurement
Prediction
z(kk1)=H(k)x(kk1)
Innovation
(k)=z(k)z(kk1)
Updated State
Estimate
x(kk)=x(kk1)+W(k) (k)
State Error Covariance at t
k1
P(k1k1)
State Prediction
Covariance
P(kk1)=
F(k)P(k1k1)F(k)+Q(K)
Innovation Covariance
S(k)=H(k)P(Kk1)H'(k)+R(k)
Filter Gain
W(k)=P(kk1)H'(k)S
1
(k)
Updated State
Covariance
P(kk)=P(kk1)W(k)S(k)W'(k)
State Covariance
Computation
Estimation
of State
True State
Control at t
k
u(k)
State Transition
x(k)=F(k)x(k1)
+G(k)u(k)+v(k)
Measurement at t
k
z(k)=H(k)x(k)+w(k)
Figure 16: Block diagram of the Kalman lter cycle (after Barshalom and Fortmann 1988
[5])
observation model in generating an update. The update stage of the lter is clearly
linear, with a weight W(k) being associated with the observation z(k) and a weight
1 W(k)H(k) being associated with the prediction x(k [ k 1). The Kalman lter also
provides a propagation equation for the covariance in the prediction and estimate.
Example 17
It is straightforward to build a stateestimator for the target and observation model
of example Example 16. The Kalman lter Equations 8385 are implemented with F(k),
Q(k) as dened in Equation 76 and with H(k) and R(k) as dened in Equation 77.
Initial conditions for x(0 [ 0) are determined from the rst few observations and with
P(0 [ 0) = 10Q(k) (see [20] for details). The results of the lter implementation are
shown in Figure 17. The Figure shows results for the x component of the track. It should
be noted that the estimate always lies between the observation and prediction (it is a
weighted sum of these terms). Note also the error (between true and estimated) velocity
BAE SYSTEMS Proprietary Information, ASCTA use only 54
and heading is zero mean and white indicating a good match between model and estimator.
3.1.3 The Innovation
A prediction can be made as to what observation will be made at a time k based on the
observations that have been made up to time k 1 by simply taking expectations of the
observation Equation 72 conditioned on previous observations:
z(k [ k 1)
= Ez(k) [ Z
k1
= EH(k)x(k) +W(k) [ Z
k1
r
= 0.1. Observations are generated with random time intervals uniformly distributed
between 0 and 2 seconds (an average of 1 second). Figure 19(a) shows the computed
position estimate and prediction standard deviations . Clearly, these covariances are not
constant, unsurprisingly, the variance in position estimate grows in periods where there
is a long time interval between observations and reduces when there is a small interval
between observations. Figure 19(b) shows the lter innovations and associated computed
standard deviations. Again, these are not constant, however they do still satisfy criteria
for whiteness with 95% of innovations falling within their corresponding 2 bounds.
BAE SYSTEMS Proprietary Information, ASCTA use only 59
300 305 310 315 320 325 330 335 340 345 350
9700
9750
9800
9850
9900
9950
10000
10050
10100
10150
10200
Time (s)
P
o
s
it
io
n
(
m
)
True Target Track and Observations of Position
Target True Position
Target Observations
(a)
300 305 310 315 320 325 330 335 340 345 350
9700
9750
9800
9850
9900
9950
10000
10050
10100
10150
10200
Time (s)
P
o
s
it
io
n
(
m
)
Predicted, Observed and Estimated Target Position
Predicted Target Position
Estimated Target Position
Position Observation
(b)
300 305 310 315 320 325 330 335 340 345 350
40
30
20
10
0
10
20
30
40
Time(s)
I
n
n
o
v
a
t
io
n
(
m
)
X Innovations and Innovation Standard Deviation
Innovation
Innovation Standard Deviation
(c)
300 305 310 315 320 325 330 335 340 345 350
5
6
7
8
9
10
11
12
13
Time(s)
P
o
s
it
io
n
E
r
r
o
r
(
m
)
Estimated and Predicted Position Standard Deviations
Predicted Postion SD
Estimated Position SD
(d)
Figure 19: Estimated target track for the linear constant velocity particle model with
asynchronous observations and updates. A timesection shown of: (a) True state and
asynchronous observations; (b) State predictions and estimates, together with observa
tions; (c) Innovations and innovation standard deviations; (d) State prediction and state
estimate standard deviations (estimated errors).
BAE SYSTEMS Proprietary Information, ASCTA use only 60
To deal with delayed observations is to maintain two estimates, one associated with
the true time at which the last observation was obtained, and the second a prediction,
based on this last estimate, which describes the state at the current time. When a new,
delayed, observation is made, the current prediction is discarded and a new prediction up
to the time of this delayed observation is computed, based on the estimate at the time
of the previous delayed observation. This is then combined with the new observation to
produce a new estimate which is itself predicted forward to the current time. Let t
c
be the
current time at which a new delayed observation is made available, let t
p
be the previous
time a delayed observation was made available and let t
c
> t
k
> t
k1
, and t
c
> t
p
> t
k1
.
We assume that we already have an estimate x(t
k1
[ t
k1
) and an estimate (strictly a
prediction) x(t
p
[ t
k1
) and that we acquire an observation z(t
k
) taken at time t
k
. We start
by simply discarding the estimate x(t
p
[ t
k1
), and generating a new prediction x(t
k
[ t
k1
)
and prediction covariance P(t
k
[ t
k1
) from Equations 93 and 94. These together with the
delayed observation are used to compute a new estimate x(t
k
[ t
k
), at the time the delayed
observation was made, from Equations 97, 98, and 99. This estimate is then predicted
forward to the current time to produce an estimate x(t
c
[ t
k
) and its associated covariance
P(t
c
[ t
k
) according to
x(t
c
[ t
k
) = F(t
p
) x(t
k
[ t
k
) +B(t
p
)u(t
p
), (100)
P(t
c
[ t
k
) = F(t
p
)P(t
k
[ t
k
)F
T
(t
p
) +G(t
p
)Q(t
p
)G
T
(t
p
). (101)
Both estimates x(t
k
[ t
k
) and x(t
p
[ t
k
) together with their associated covariances should
be maintained for the next observations.
It should be clear that if the observations are delayed, then the estimate provided at
the current time will not be as good as the estimates obtained when the observations are
obtained immediately. This is to be expected because the additional prediction required
injects additional process noise into the state estimate.
It is also useful to note that the same techniques can be used to produce estimates
for any time in the future as well as simply the current time. This is sometimes use
ful in providing advance predictions that are to be used to synchronize with incoming
observations.
Asequent data occurs when the observations made are delayed in such a way that
they arrive at the lter for processing out of timeorder. Although this does not often
happen in singlesensor systems, it is a common problem in multisensor systems where
the preprocessing and communication delays may dier substantially between dierent
sensors. The essential problem here is that the gain matrix with and without the delayed
observation will be dierent and the previous estimate, corresponding to the time at
which the delayed observation was taken, cannot easily be unwrapped from the current
estimate. With the tools we have so far developed, there is no simple way of dealing with
this problem other than by remembering past estimates and recomputing the current
estimate every time an outoforder observation is obtained. However, a solution to this
problem is possible using the inversecovariance lter which we will introduce latter in
this chapter.
BAE SYSTEMS Proprietary Information, ASCTA use only 61
3.1.5 The Extended Kalman Filter
The extended Kalman lter (EKF) is a form of the Kalman lter that can be employed
when the state model and/or the observation model are nonlinear. The EKF is briey
described in this section.
The state models considered by the EKF are described in statespace notation by a
rst order nonlinear vector dierential equation or state model of the form
x(t) = f[x(t), u(t), v(t), t], (102)
where
x(t) 1
n
is the state of interest,
u(t) 1
r
is a known control input,
f[, , ] a mapping of state and control input to state velocity, and
v(t) a random vector describing both dynamic driving noise and uncertainties in the
state model itself (v(t) is often assumed additive).
The observation models considered by the EKF are described in statespace notation by
a nonlinear vector function in the form
z(t) = h[x(t), u(t), w(t), t] (103)
where
z(t) 1
m
is the observation made at time t,
h[, , ] is a mapping of state and control input to observations, and
w(t) a random vector describing both measurement corruption noise and uncertainties
in the measurement model itself (w(t) is often assumed additive).
The EKF, like the Kalman lter, is almost always implemented in discretetime. To do
this, a discrete form of Equations 102 and 103 are required. First, a discretetime set
t = t
0
, t
1
, t
k
, is dened. Equation 103 can then be written in discrete time as
z(t
k
) = h[x(t
k
), u(t
k
), w(t
k
), t
k
], t
k
t (104)
where z(t
k
), x(t
k
) and w(t
k
) are the discretetime observation, state and noise vectors
evaluated at the discrete time instant t
k
. The discretetime form of the state equation
requires integration of Equation 102 over the interval (t
k
, t
k1
) as
x(t
k
) = x(t
k1
) +
_
t
k
t
k1
f[x(), u(), v(), ]d. (105)
BAE SYSTEMS Proprietary Information, ASCTA use only 62
In practice, this integration is normally solved using a simple Euler (backward dierence)
approximation as
x(t
k
) = x(t
k1
) + T
k
f[x(t
k1
), u(t
k1
), v(t
k1
), t
k1
], (106)
where T
k
= t
k
t
k1
. As with the Kalman lter, when the sample interval is constant,
time is indexed by k and Equations 106 and 104 are written as
x(k) = x(k 1) + Tf[x(k 1), u(k 1), v(k 1), k], (107)
and
z(k) = h[x(k), u(k), w(k), k]. (108)
To apply the Kalman lter algorithm to estimation problems characterised by non
linear state and observation models, perturbation methods are used to linearise true non
linear system models around some nominal state trajectory to yield a model which is
itself linear in the error. Given a nonlinear system model in the form of Equation 102,
a nominal state trajectory is described using the same process model (assuming v(t) is
zero mean),
x
n
(t) = f[x
n
(t), u
n
(t), t]. (109)
Then, Equation 102 is expanded about this nominal trajectory as a Taylor series;
x(t) = f[x
n
(t), u
n
(t), t]
+f
x
(t)x(t) +O
_
x(t)
2
_
+f
u
(t)u(t) +O
_
u(t)
2
_
+f
v
(t)v(t), (110)
where
f
x
(t)
=
f
x
x(t)=x
n
(t)
u(t)=u
n
(t)
, f
u
(t)
=
f
u
x(t)=x
n
(t)
u(t)=u
n
(t)
, f
v
(t)
=
f
v
x(t)=x
n
(t)
u(t)=u
n
(t)
(111)
and
x(t)
= (x(t) x
n
(t)) , u(t)
= (u(t) u
n
(t)) . (112)
Subtracting Equation 109 from Equation 110 provides a linear error model in the form
x(t) = f
x
(t)x(t) +f
u
(t)u(t) +f
v
(t)v(t). (113)
Identifying F(t) = f
x
(t), B(t) = f
u
(t), and G(t) = f
v
(t), Equation 113, is now in
the same form as Equation 63 and may be solved for the perturbed state vector x(t)
in closed form through Equation 66, yielding a linear discretetime equation for error
propagation. Clearly this approximation is only valid when terms of second order and
higher are small enough to be neglected; when the true and nominal trajectories are close
and f() is suitably smooth. With judicious design of the estimation algorithm this can
BAE SYSTEMS Proprietary Information, ASCTA use only 63
be achieved surprisingly often. It is also possible to retain or approximate higherorder
terms from Equation 110 and so improve the validity of the approximation.
Similar arguments can be applied to linearise Equation 64 to provide an observation
error equation in the form
z(t) = h
x
(t)x(t) +f
h
(t)w(t). (114)
Identifying H(t) = f
x
(t), and D(t) = f
w
(t), Equation 114, is now in the same form as
Equation 64.
The discretetime extended Kalman lter algorithm can now be stated. With appro
priate identication of discrete time states and observations, the state model is written
as
x(k) = f (x(k 1), u(k), v(k), k) , (115)
and the observation model as
z(k) = h(x(k), w(k)) . (116)
Like the Kalman lter, it is assumed that the noises v(k) and w(k) are all Gaussian, tem
porally uncorrelated and zeromean with known variance as dened in Equations 7375.
The EKF aims to minimise meansquared error and therefore compute an approximation
to the conditional mean. It is assumed therefore that an estimate of the state at time
k 1 is available which is approximately equal to the conditional mean,
x(k 1 [ k 1) Ex(k 1) [ Z
k1
. (117)
The extended Kalman lter algorithm will now be stated without proof. Detailed deriva
tions may be found in any number of books on the subject. The principle stages in the
derivation of the EKF follow directly from those of the linear Kalman lter with additional
step that the process and observation models are linearised as a Taylors series about the
estimate and prediction respectively. The algorithm has two stages:
Prediction: A prediction x(k [ k 1) of the state at time k and its covariance P(k [ k 1)
is computed according to
x(k [ k 1) = f ( x(k 1 [ k 1), u(k)) (118)
P(k [ k 1) = f
x
(k)P(k 1 [ k 1)
T
f
x
(k) +f
v
(k)Q(k)
T
f
v
(k) (119)
Update: At time k an observation z(k) is made and the updated estimate x(k [ k) of
the state x(k), together with the updated estimate covariance P(k [ k) is computed from
the state prediction and observation according to
x(k [ k) = x(k [ k 1) +W(k) [z(k) h( x(k [ k 1))] (120)
P(k [ k) = P(k [ k 1) W(k)S(k)W
T
(k) (121)
BAE SYSTEMS Proprietary Information, ASCTA use only 64
where
W(k) = P(k [ k 1)
T
h
x
(k)S
1
(k) (122)
and
S(k) = h
x
(k)P(k [ k 1)
T
h
x
(k) +h
w
(k)R(k)
T
h
w
(k). (123)
and where the Jacobian f
(k) is
evaluated at and x(k) = x(k [ k 1).
A comparison of Equations 8192 with Equations 118123 makes it clear that the
extended Kalman lter algorithm is very similar to the linear Kalman lter algorithm,
with the substitutions F(k) f
x
(k) and H(k) h
x
(k) being made in the equations
for the variance and gain propagation. Thus, the extended Kalman lter is, in eect, a
linear estimator for a state error which is described by a linear equation and which is
being observed according to a linear equation of the form of Equation 72.
The extended Kalman lter works in much the same way as the linear Kalman lter
with some notable caveats:
The Jacobians f
x
(k) and h
x
(k) are typically not constant, being functions of
both state and timestep. This means that unlike the linear lter, the covariances
and gain matrix must be computed online as estimates and predictions are made
available, and will not in general tend to constant values. This signicantly increase
the amount of computation which must be performed online by the algorithm.
As the linearised model is derived by perturbing the true state and observation
models around a predicted or nominal trajectory, great care must be taken to ensure
that these predictions are always close enough to the true state that second order
terms in the linearisation are indeed insignicant. If the nominal trajectory is too
far away from the true trajectory then the true covariance will be much larger than
the estimated covariance and the lter will become poorly matched. In extreme
cases the lter may also become unstable.
The extended Kalman lter employs a linearised model which must be computed
from an approximate knowledge of the state. Unlike the linear algorithm, this means
that the lter must be accurately initialized at the start of operation to ensure that
the linearised models obtained are valid. If this is not done, the estimates computed
by the lter will simply be meaningless.
3.1.6 The Covariance Intersect (CI) Algorithm
The Covariance Intersection (CI) algorithm provides a solution to the problem of com
bining two random vectors in the case where the correlation between these vectors is
unknown [44]. The algorithm is based on a geometric interpretation of the normal covari
ance or information matrices involved in ltering problems. Figure 20 demonstrates the
principle. Consider two estimates a and b with covariances P
a
and P
b
respectively. For
simplicity and future compatability, denote Y
a
= P
1
a
and Y
a
= P
1
a
. Figure 20(a) shows
BAE SYSTEMS Proprietary Information, ASCTA use only 65
4.76 4.78 4.8 4.82 4.84 4.86 4.88
x 10
4
6.02
6.04
6.06
6.08
6.1
6.12
6.14
x 10
4
Xvariance (m
2
)
Y
v
a
r
ia
n
c
e
(
m
2
)
Covariance 1
Covariance 2
Updated Covariance
(a)
4.76 4.78 4.8 4.82 4.84 4.86 4.88
x 10
4
6.02
6.04
6.06
6.08
6.1
6.12
6.14
x 10
4
Xvariance (m
2
)
Y
v
a
r
ia
n
c
e
(
m
2
)
Covariance 1
Covariance 2
(b)
4.76 4.78 4.8 4.82 4.84 4.86 4.88
x 10
4
6.02
6.04
6.06
6.08
6.1
6.12
6.14
x 10
4
Xvariance (m
2
)
Y
v
a
r
ia
n
c
e
(
m
2
)
Covariance 1
Covariance 2
Updated Covariance
(c)
5.08 5.1 5.12 5.14 5.16 5.18
x 10
4
6.44
6.46
6.48
6.5
6.52
6.54
x 10
4
Xvariance (m
2
)
Y
v
a
r
ia
n
c
e
(
m
2
)
Covariance 1
Covariance 2
Updated Covariance
(d)
Figure 20: Operation of the covariance intersection (CI) algorithm: (a) The initial two es
timate covariances and the updated covariance computed by the Kalman lter algorithm;
(b) The possible updated covariances computed from a convex combination of the two
initial covariances parameterized by the scalar ; (c) The updated covariance as the con
vex combination of the two initial covariances which maximises information (minimises
enclosed volume); (d) The case where one initial covariance dominates the other.
BAE SYSTEMS Proprietary Information, ASCTA use only 66
two (equiprobable) ellipsoids generated from these covariance matrices together with the
combined estimate covariance ellipse calculated by the Kalman lter as:
P
1
c
= P
1
a
+P
1
b
or equivalently
Y
c
= Y
a
+Y
b
.
The assumption in the Kalman lter is that the two estimates are independent. If this is
not the case then the updated covariance computed by the Kalman lter will be overly
optimistic and the lter will become inconsistent.
Instead, the CI algorithm computes an updated covariance matrix as a convex com
bination of the two initial covariance matrices in the form
Y
c
= (1 )Y
a
+Y
b
(124)
y
c
= (1 )y
a
+y
b
. (125)
Figure 20(b) shows these dierent ellipses as varies between 0 and 1. This family of
ellipses all enclose the intersection of the two initial covariance ellipsoids. Any of these
ellipses would provide a consistent estimate of the updated covariance. However, a value
of is normally chosen which minimizes the volume of the updated covariance (this is
equivalent to minimizing the determinant of Y
c
). This minimum is shown in Figure 20(c).
It is clear the the CI algorithm can be quite conservative. Indeed, in many cases, the CI
algorithm will set to an extrema of 1 or 0 when one of the initial covariances dominates
the other. This results in no new information being fused as is shown in Figure 20(d).
This must be weighed against the fact that the CI algorithm will guarantee consistency
for cases when the correlation between two estimates is unknown.
3.2 The MultiSensor Kalman Filter
Many of the techniques developed for single sensor Kalman lters can be applied directly
to multisensor estimation and tracking problems. In principle, a group of sensors can
be considered as a single sensor with a large and possibly complex observation model. In
this case the Kalman lter algorithm is directly applicable to the multisensor estimation
problem. However, as will be seen, this approach is practically limited to relatively small
numbers of sensors.
A second approach is to consider each observation made by each sensor as a separate
and independent realization, made according to a specic observation model, which can
be incorporate into the estimate in a sequential manner. Again, singlesensor estimation
techniques, applied sequentially, can be applied to this formulation of the multisensor
estimation problem. However, as will be seen, this approach requires that a new prediction
and gain matrix be calculated for each observation from each sensor at every timestep,
and so is computationally very expensive.
A third approach is to explicitly derive equations for integrating multiple observations
made at the same time into a common state estimate. Starting from the formulation of
BAE SYSTEMS Proprietary Information, ASCTA use only 67
the multisensor Kalman lter algorithm, employing a single model for a group of sensors,
a set of recursive equations for integrating individual sensor observations can be derived.
As will be seen, these equations are more naturally expressed in terms of information
rather than state and covariance.
The systems considered to this point are all centralized; the observations made by
sensors are reported back to a central processing unit in a raw form where they are
processed by a single algorithm in much the same way as single sensor systems. It is also
possible to formulate the multisensor estimation problem in terms of a number of local
sensor lters, each generating state estimates, which are subsequently communicated in
processed form back to a central fusion centre. This distributed processing structure has
a number of advantages in terms of modularity of the resulting architecture. However,
the algorithms required to fuse estimate or track information at the central site can be
quite complex.
In this section, the multisensor estimation problem is rst dened in terms of a set of
observation models and a single common process model. Each of the three centralised pro
cessing algorithms described above will be developed and compared. techniques described
above; deriving appropriate equations and developing simple examples. The following sec
tion will then show how these multisensor estimation algorithms are applied in simple
tracking problems.
Example 20
This example introduces a fundamental tracking problem which is further developed in
the remainder of this section. The problem consists of the tracking of a number of targets
from a number of tracking stations. The simulated target models and observations are
nonlinear, while the tracking algorithms used at the sensor sites are linear. This is fairly
typical of actual tracking situations.
The targets to be tracked are modeled as 2dimensional platforms with controllable
heading and velocity. The continuous model is given by:
_
_
x(t)
y(t)
(t)
_
_ =
_
_
V (t) cos( +)
V (t) sin( +)
V (t)
sin()
_
_
where (x(t), y(t)) is the target position, (t) is the target orientation, V (t) and (t) are the
platform velocity and heading, and is a constant minimum instantaneous turn radius
for the target. Then x(t) = [x(t), y(t), (t)]
T
is dened as the state of the target and
u(t) = [V (t), (t)]
T
as the target input.
This model can be converted into a discrete time model using a constant (Euler) inte
gration rule over the (asynchronous) sample interval T
k
= t
k
t
k1
as
_
_
x(k)
y(k)
(k)
_
_ =
_
_
x(k 1) + TV (k) cos((k 1) +(k))
y(k 1) + TV (k) sin((k 1) +(k))
T(k 1) +
V (k)
sin((k))
_
_
BAE SYSTEMS Proprietary Information, ASCTA use only 68
which is in the standard form x(k) = f(x(k 1), u(k)). For the purposes of simulation,
the model is assumed driven by Brownian models for velocity and heading,
V (t) = V (t) +v
V
(t), (t) = (t) +v
(t),
where v(t) = [v
V
(t), v
(t)]
T
is a zero mean white driving sequence with strength
Ev(t)v
T
(t) =
_
2
v
0
0
2
_
The discretetime model for this is simply
V (k) = V (k 1) + T
k
v
V
(k), (k) = (k 1) + T
k
v
(k).
A group of typical trajectories generated by this model is shown in Figure 21(a).
The targets are observed using sensors (tracking stations) i = 1, , N whose location
and pointing angles T
i
(t) = [X
i
(t), Y
i
(t),
i
(t)]
T
are known. This does not preclude the
sensors from actually being in motion. Each sensor site i is assumed to make range and
bearing observations to the j
th
targets as
_
z
r
ij
(k)
z
ij
(k)
_
=
_
_
_
(x
j
(k) X
i
(k))
2
+ (y
j
(k) Y
i
(k))
2
arctan
_
y
j
(k)Y
i
(k)
x
j
(k)X
i
(k)
_
i
(k)
_
_
+
_
w
r
ij
(k)
w
ij
(k)
_
,
where the random vector w
ij
(k) = [r
r
ij
(k), r
ij
(k)]
T
describes the noise in the observation
process due to both modeling errors and uncertainty in observation. Observation noise
errors are taken to be zero mean and white with constant variance
Ew
ij
(k)w
T
ij
(k) =
_
2
r
0
0
2
_
.
A typical set of observations generated by this model for a track is shown in Figure 21(b).
3.2.1 Observation Models
Figure 22 shows the centralised data fusion architecture developed in the following three
sections. A common model of the true state is provided in the usual linear discretetime
form;
x(k) = F(k)x(k 1) +G(k)u(k) +v(k), (126)
where x() is the state of interest, F(k) is the state transition matrix, G(k)the control
input model, u(k) the control input vector, and v(k) a random vector describing model
and process uncertainty, assumed zero mean and temporally uncorrelated;
Ev(k) = 0, k,
Ev(i)v
T
(j) =
ij
Q(i).
(127)
BAE SYSTEMS Proprietary Information, ASCTA use only 69
3 4 5 6 7 8
x 10
4
3.5
4
4.5
5
5.5
6
6.5
7
7.5
8
x 10
4
True Target Motions
xposition (m)
y
p
o
s
i
t
i
o
n
(
m
)
Target True Position
(a)
6 6.2 6.4 6.6 6.8 7 7.2 7.4 7.6 7.8
x 10
4
5
5.2
5.4
5.6
5.8
6
6.2
6.4
x 10
4
True Target Motions
xposition (m)
y
p
o
s
i
t
i
o
n
(
m
)
Target True Position
Tracking Stations
Target Observations
(b)
Figure 21: Typical tracks and observations generated by the standard multitarget track
ing example of Example 20: (a) true xy tracks; (b) detail of track observations.
BAE SYSTEMS Proprietary Information, ASCTA use only 70
w
1
H
1
Sensor 1
+
+
w
s
H
s
Sensor s
+
+
w
S
H
S
Sensor S
+
+
...... ......
x(k)=Fx(k1)+ Gu(k)+v(k)
z
1
z
S
z
s
Fusion Centre
Generating x(kk)
Figure 22: The structure of a centralised data fusion architecture. Each sensor
has a dierent observation model but observes a common process. Observa
tions are sent in unprocessed form to a central fusion center that combines
observations to provide an estimate of the common state.
It is important to emphasise that, because all sensors are observing the same state (there
would be little point in the data fusion problem otherwise) this process model must be
common to all sensors.
Observations of the state of this system are made synchronously by a number of
dierent sensors according to a set of linear observation models in the form
z
s
(k) = H
s
(k)x(k) +w
s
(k), s = 1, , S. (128)
where z
s
(k) is the observation made at time k by sensor s of a common state x(k)
according to the observation model H
s
(k) in additive noise w
s
(k). The case in which
the observations made by the sensors are asynchronous, in which dierent sensors make
observations at dierent rates, can be dealt with by using the observation model H
s
(t
k
).
It is assumed that the observation noise models w
s
(k) are all zero mean, uncorrelated
between sensors and also temporally uncorrelated;
Ew
s
(k) = 0, s = 1, , S, k
Ew
p
(i)w
q
(j) =
ij
pq
R
p
(i).
(129)
BAE SYSTEMS Proprietary Information, ASCTA use only 71
It is also assumed that the process and observation noises are uncorrelated
Ev(i)w
T
s
(j) = 0, i, j, s. (130)
This assumption is not absolutely necessary. It is relatively simple, but algebraically
complex, to include a term accounting for correlated process and observation errors (see
[12] Chapter 7, or [28].)
3.2.2 The GroupSensor Method
The simplest way of dealing with a multisensor estimation problem is to combine all
observations and observation models in to a single composite group sensor and then
to deal with the estimation problem using an identical algorithm to that employed in
singlesensor systems. Dening a composite observation vector by
z(k)
=
_
z
T
1
(k), , z
T
S
(k)
_
T
, (131)
and a composite observation model by
H(k)
=
_
H
T
1
(k), , H
T
s
(k)]
_
T
, (132)
with
w(k)
=
_
w
T
1
(k), , w
T
s
(k)
_
T
, (133)
where from Equation 129
R(k) = Ew(k)w
T
(k)
= E
_
w
T
1
(k), , w
T
s
(k)
_
T
_
w
T
1
(k), , w
T
s
(k)
_
= blockdiagR
1
(k), , R
s
(k),
(134)
the observation noise covariance is blockdiagonal with blocks equal to the individual
sensor observation noise covariance matrices. The set of observation equations dened
by Equation 128 may now be rewritten as a single group observation model in standard
form
z(k) = H(k)x(k) +w(k). (135)
With a process model described by Equation 126 and a group observation model dened
by Equation 135, estimates of state can in principle be computed using the standard
Kalman lter algorithm given by Equations 90, 91, and 92.
Example 21
Consider again the tracking of a particle moving with constant velocity with process
model as dened in Example 16. Suppose we have two sensors, the rst observing the
BAE SYSTEMS Proprietary Information, ASCTA use only 72
position and the second observing the velocity of the particle. The two observation models
will be
z
1
= [ 1 0 ]
_
x(k)
x(k)
_
+w
1
(k), Ew
2
1
(k) =
2
r
1
,
and
z
2
= [ 0 1 ]
_
x(k)
x(k)
_
+w
2
(k), Ew
2
2
(k) =
2
r
2
.
The composite group sensor model is simply given by
_
z
1
z
2
_
=
_
1 0
0 1
_ _
x(k)
x(k)
_
+
_
w
1
w
2
_
, E
_
w
1
w
2
_
[ w
1
w
2
] =
_
2
r
1
0
0
2
r
2
_
.
We could add a third sensor making additional measurements of position according to the
observation model
z
3
= [ 1 0 ]
_
x(k)
x(k)
_
+w
3
(k), Ew
2
3
(k) =
2
r
3
,
in which case the new composite group sensor model will be given by
_
_
z
1
z
2
z
3
_
_ =
_
_
1 0
0 1
1 0
_
_
_
x(k)
x(k)
_
+
_
_
w
1
w
2
w
3
_
_,
E
_
_
w
1
w
2
w
3
_
_ [ w
1
w
2
w
3
] =
_
2
r
1
0 0
0
2
r
2
0
0 0
2
r
3
_
_.
It should be clear that there is no objection in principle to incorporating as many sensors
as desired in this formulation of the multisensor estimation problem.
The prediction phase of the multisensor Kalman lter algorithm makes no reference to
the observations that are made and so is identical in every respect to the prediction phase
of the singlesensor lter. However, the update phase of the cycle, which incorporates
measurement information from sensors, will clearly be aected by an increase in the
number of sensors. Specically, if we have a state vector x(k) of dimension n and S
sensors each with an observation vector z
s
(k), s = 1, , S of dimension m
s
together
with an observation model H
s
(k) of dimension m
s
n then the group observation vector
z(k) will have dimension m =
S
s=1
m
s
and the group observation model H(k) will have
dimension m n. The consequence of this lies in Equations 90, 91 and 92. Clearly
the groupsensor innovation (k)
=
_
T
1
(k), ,
T
m
(k)
_
T
will now have dimension m, and
the groupsensor innovation covariance matrix S(k) will have dimension mm. As the
number of sensors incorporated in the group sensor model increases, so does the dimension
of the innovation vector and innovation covariance matrix. This is a problem because the
inverse of the innovation covariance is required to compute the groupsensor gain matrix
BAE SYSTEMS Proprietary Information, ASCTA use only 73
W(k) in Equation 92. It is well known that the calculation of a matrix inverse increase
in proportion to the square of its dimension.
For a small number of sensors and so for a relatively small innovation dimension, the
groupsensor approach to multisensor estimation may be the most practical to implement.
However, as the number of sensors increases, the value of this monolithic approach to the
data fusion problem becomes limited.
Example 22
It is common practice to use linear models to track targets that are clearly not linear,
particularly in multipletarget tracking problems. Consider again the tracks and obser
vations generated by Example 20. The range and bearing observation vector z
ij
(k) =
[z
r
ij
(k), z
ij
(k)]
T
can be converted into an equivalent observation vector in absolute carte
sian coordinates as
_
z
x
ij
(k)
z
y
ij
(k)
_
=
_
X
i
(k) +z
r
ij
(k) cos z
ij
(k)
Y
i
(k) +z
r
ij
(k) sin z
ij
(k)
_
.
The observation covariance matrix can also be converted into absolute cartesian coordi
nates using the relationship
R
xy
(k) =
_
2
xx
xy
xy
2
yy
_
= Rot(z
ij
(k))
_
2
r
0
0 (z
r
ij
(k))
2
_
Rot
T
(z
ij
(k))
where Rot() is the rotation matrix
Rot() =
_
cos sin
sin cos
_
.
Note now that observation variance is strongly range dependent and is lined up with the
sensor boresight.
Once the observation vector has been converted into a global cartesian coordinate frame,
a linear lter can be used to estimate a linear target track. The twodimensional particle
model of Example 16 with process model dened in Equation 76 and with an observation
model (now not constant) dened in Equation 77, can be used in the lter dened by
Equations 8385 in Example 17.
Figure 23 shows the results of tracking four targets from two (stationary) sensor sites.
First note that the state variances and innovation variances are not constant because the
observation variances are strongly dependent on the relative position of observer and tar
get. The asynchronous nature of the observations also contributes to this nonconstancy.
It can be observed in Figure 23(c) and (d) the rising and falling as variance values as the
target shown comes closer to and then further away from the tracking sites.
The algorithm implemented here is equivalent to the group sensor method, but the
results will also be the same for the sequential and inverse covariance algorithms.
BAE SYSTEMS Proprietary Information, ASCTA use only 74
4.5 5 5.5 6 6.5 7 7.5 8
x 10
4
3
3.5
4
4.5
5
5.5
6
6.5
7
7.5
x 10
4
True Target Motions
xposition (m)
y
p
o
s
it
io
n
(
m
)
Target True Position
Tracking Stations
Target Observations
(a)
6.05 6.1 6.15 6.2 6.25 6.3 6.35 6.4
x 10
4
5.5
5.6
5.7
5.8
5.9
6
x 10
4
Estimated Tracks
xposition (m)
y
p
o
s
it
io
n
(
m
)
Target True Position
Tracking Stations
Target Observations
Estimated Tracks
(b)
100 150 200 250 300 350 400
800
600
400
200
0
200
400
600
800
xinnovation for Track 1 from Tracking Site 1
Time (s)
I
n
n
o
v
a
t
io
n
(
m
)
(c)
50 100 150 200 250 300 350 400 450
200
150
100
50
0
50
100
150
xestimate error for Track 1 from all Tracking Sites
Time (s)
E
r
r
o
r
(
m
)
(d)
Figure 23: A multiple sensor multiple target tracking example with four targets and two
tracking stations: (a) True state and asynchronous observations; (b) Detail of true states,
track estimates, and observations; (c) Innovations and innovation standard deviations for
a particular track and tracking station; (d) Track position estimates from all sites together
with standard deviations (estimated errors).
BAE SYSTEMS Proprietary Information, ASCTA use only 75
3.2.3 The SequentialSensor Method
A second approach to the multisensor estimation problem is to consider each sensor obser
vation as an independent, sequential update to the state estimate and for each observation
to compute an appropriate prediction and gain matrix. In contrast to the groupsensor
approach in which a single sensor model was constructed by combining all sensor mod
els, the sequential update approach considers each sensor model individually, one by one.
This means that the dimension of the innovation and innovation covariance matrix at
each update stage remains the same size as their singlesensor equivalents at the cost of
computing a new gain matrix for each observation from each sensor.
The description of the state to be estimated is assumed in the form of Equation 126.
Observations are made of this common state by S sensors according to Equation 128. It
is assumed that every sensor takes an observation synchronously at every time step. The
case of asynchronous observations is straightforward in this sequentialsensor algorithm.
It is assumed that it is possible to consider the observations made by the sensors at
any one timestep in a specic but otherwise arbitrary order so that the observation z
p
(k),
1 p S, from the p
th
sensor will be before the observation z
p+1
(k) from the (p + 1)
th
.
The set of observations made by the rst p sensors at time k is denoted by (caligraphic
notation is used to denote sets across sensors compared to boldface notation for sets
associated with a single sensor)
Z
p
(k)
= z
1
(k), , z
p
(k) , (136)
so that at every timestep the observation set Z
S
(k) is available to construct a new state
estimate. The set of all observations made by the p
th
sensor up to time k will be denoted
by
Z
k
p
= z
p
(1), , z
p
(k) , (137)
and the set of all observations made by the rst p sensors up to time k by
Z
k
p
= Z
k
1
, , Z
k
p
, p S. (138)
Of particular interest is the set of observations consisting of all observations made by the
rst p sensors up to a time i and all observations made by the rst q sensors up to a time
j
Z
i,j
p,q
= Z
i
p
Z
j
q
= Z
i
1
, , Z
i
p
, Z
j
p+1
, , Z
j
q
, p < q, i j. (139)
and in particular the set consisting of all observations made by all sensors up to a time
k 1 together with the observations made by the rst p sensors up to time k
Z
k,k1
p,S
= Z
k
p
Z
k1
S
= Z
k
1
, , Z
k
p
, Z
k1
p+1
, , Z
k1
S
= Z
p
(k) Z
k1
S
, p < S.
(140)
BAE SYSTEMS Proprietary Information, ASCTA use only 76
The estimate constructed at each timestep on the basis of all observations from all sensors
up to time k1 and on the observations made by the rst p sensors up to time k is dened
by
x(k [ k, p) = Ex(k) [ Z
k,k1
p,S
, (141)
where in particular
x(k [ k, 0) = x(k [ k 1) = Ex(k) [ Z
k1
S
, (142)
is the prediction of the state at time k before any observations are made and
x(k [ k, S) = x(k [ k) = Ex(k) [ Z
k
S
, (143)
is the estimate at time k based on the observations made by all sensors.
It is assumed that Equations 127, 129, and 130 hold. For the rst observation con
sidered at any one timestep, the prediction stage for the sequentialsensor estimator is
very similar to the prediction stage for the singlesensor estimator. The state prediction
is simply
x(k [ k, 0) = x(k [ k 1)
= F(k) x(k 1 [ k 1) +G(k)u(k)
= F(k) x(k 1 [ k 1, S) +G(k)u(k),
(144)
with corresponding covariance
P(k [ k, 0) = P(k [ k 1)
= F(k)P(k 1 [ k 1)F
T
(k) +Q(k)
= F(k)P(k 1 [ k 1, S)F
T
(k) +Q(k).
(145)
The new estimate found by integrating the observation z
1
(k) made by the rst sensor at
time k is then simply given by
x(k [ k, 1) = x(k [ k, 0) +W
1
(k) [z
1
(k) H
1
(k) x(k [ k, 0)] (146)
with covariance
P(k [ k, 1) = P(k [ k, 0) W
1
(k)S
1
(k)W
T
1
(k), (147)
where
W
1
(k) = P(k [ k, 0)H
T
1
(k)S
1
1
(k) (148)
and
S
1
(k) = H
1
(k)P(k [ k, 0)H
T
1
(k) +R
1
(k) (149)
To now integrate the observation made by the second sensor we need to employ the
estimate x(k [ k, 1) to generate a prediction. However, if the observations made by the
BAE SYSTEMS Proprietary Information, ASCTA use only 77
sensors are assumed synchronous, the state does not evolve between successive sensor
readings from the same timestep and so F(k) = 1 and Q(k) = 0 (this is not true in
the asynchronous case). This means that the estimate x(k [ k, 1) is itself the prediction
for the next update. In general, the estimate of the state x(k [ k, p) at time k based on
observations made by all sensors up to time k 1 and by the rst p sensors up to time k
can be computed from the estimate x(k [ k, p 1) as
x(k [ k, p) = x(k [ k, p 1) +W
p
(k) [z
p
(k) H
p
(k) x(k [ k, p 1)] (150)
with covariance
P(k [ k, p) = P(k [ k, p) W
p
(k)S
p
(k)W
T
p
(k), (151)
where
W
p
(k) = P(k [ k, p 1)H
T
p
(k)S
1
p
(k) (152)
and
S
p
(k) = H
p
(k)P(k [ k, p 1)H
T
p
(k) +R
p
(k) (153)
The state update at each timestep could be computed in a batchmode by explicitly
expanding Equation 150 to become
x(k [ k) =
_
S
i=1
(1 W
i
(k)H
i
(k))
_
x(k [ k 1)
+
S
i=1
_
S
j=i+1
(1 W
j
(k)H
j
(k))
_
W
i
(k)z
i
(k)
(154)
The case in which the observations are not synchronous may be dealt with in a similar
way to the singlesensor asynchronous case providing that due care is taken to generate a
proper prediction at each time each sensor records an observation.
As with the groupsensor approach to multisensor estimation, this method works
well for small numbers of sensors
7
. However the amount of computation that must be
performed increases with an increase in the number of sensors as a new gain matrix
must be computed for each observation from each sensor at each time step. Although
this increase is linear (the dimension of the innovation covariance matrix which must be
inverted does not increase), it may still become a problem when large numbers of sensors
are employed.
3.2.4 The InverseCovariance Form
Neither the groupsensor nor the sequentialsensor algorithms are of much help when the
number of sensors becomes large. In such cases it would be useful to nd an explicit set of
equations and provide an algorithm which would allow the direct integration of multiple
observations into a single composite estimate.
7
Indeed the sequentialsensor method is to be preferred when the observations are not synchronous as
in the asynchronous case a new prediction and gain matrix must be computed for each new observation
regardless.
BAE SYSTEMS Proprietary Information, ASCTA use only 78
The multisensor estimation problem would be considerably simplied if it were possi
ble to write the estimate as a simple linear combination of the innovations and prediction
in the standard Kalman lter form. Unfortunately
x(k [ k) ,= x(k [ k 1) +
S
i=1
W
i
(k) [z
i
(k) H
i
(k) x(k [ k 1)] , (155)
with
W
i
(k) = P(k [ k 1)H
T
i
(k)S
1
(k), i = 1, , S,
and
S
i
(k) = H
i
(k)P(k [ k 1)H
T
i
(k) +R
i
(k), i = 1, , S.
If the equality in Equation 155 were to hold, the groupsensor gain matrix would need to
be in blockdiagonal form. For the gainmatrix to be blockdiagonal, Equation 92 shows
that the groupsensor innovation covariance matrix must also be in blockdiagonal form.
Unfortunately, this is never the case. For example with two sensors, the groupsensor
observation vector is
z(k) = [z
T
1
(k), z
T
2
(k)]
T
,
the groupsensor observation matrix is
H(k) = [H
T
1
(k), H
T
2
(k)]
T
,
and the groupsensor observation noise covariance matrix
R(k) = blockdiagR
1
(k), R
2
(k).
The group sensor innovation covariance is thus
S(k) = H(k)P(k [ k 1)H
T
(k) +R(k)
=
_
H
1
(k)
H
2
(k)
_
P(k [ k 1) [ H
T
1
(k) H
T
2
(k) ] +
_
R
1
(k) 0
0 R
2
(k)
_
=
_
_
H
1
(k)P(k [ k 1)H
T
1
(k) +R
1
(k)
H
2
(k)P(k [ k 1)H
T
1
(k)H
1
(k)P(k [ k 1)H
T
2
(k)
H
2
(k)P(k [ k 1)H
T
2
(k) +R
2
(k)
_
_,
(156)
which is clearly not blockdiagonal; so the gain matrix
W(k) = P(k [ k 1)H
T
(k)S
1
(k)
will also not be blockdiagonal.
Fundamentally, the reason why Equation 155 may be used to integrate singlesensor
observations over time but not for integrating multiple observations at a single time
step is that the innovations are correlated
8
. This correlation is due to the fact that the
8
the innovations from one timestep to the next are uncorrelated but the innovations generated by
many sensors at a single time are correlated.
BAE SYSTEMS Proprietary Information, ASCTA use only 79
innovations at a single timestep have a common prediction. This correlation is reected
in the odiagonal terms of Equation 156 of the form H
i
(k)P(k [ k 1)H
T
j
(k).
The fact that the innovations generated by each sensor are correlated results in sig
nicant problems in the implementation of multisensor estimation algorithms. Ignoring
correlations between innovations (assuming equality in Equation 155 and constructing
an estimate from a simple weighted sum of innovations) will result in disaster. This is
because the information due to the prediction will be double counted in each update
thus implying considerably more information (and condence) than is actually the case.
The use of forgetting factors or fading memory lters are routinely used to address this
issue, although these are really no more than domainspecic hacks. It should also be
pointed out that the correlatedness of innovations in multisensor estimation exposes the
popular fallacy that it is always cheaper to communicate innovations in such algorithms.
Example 23
For a twosensor system we can explicitly evaluate the combined innovation covariance
matrix using the matrix inversion lemma and compute the appropriate gain matrices for
the innovations of both sensors. Dropping all time subscripts, the inverse innovation
covariance may be written as [30]
S
1
=
_
S
11
S
12
S
T
12
S
22
_
1
=
_
1
1
S
12
S
1
22
S
1
22
S
T
12
1
S
1
22
+S
1
22
S
T
12
1
S
12
S
1
22
_
, (157)
where
= S
11
S
12
S
1
22
S
T
12
.
Making appropriate substitutions from Equation 156, and employing the matrix inversion
lemma twice, we have
1
=
_
S
11
S
12
S
1
22
S
T
12
_
1
=
_
H
1
PH
T
1
+R
1
H
1
PH
T
2
_
H
2
PH
T
2
+R
2
_
1
H
2
PH
T
1
_
1
=
_
R
1
+H
1
_
PPH
T
2
_
H
2
PH
T
2
+R
2
_
1
H
2
P
_
H
T
1
_
1
=
_
R
1
+H
1
_
P
1
+H
T
2
R
1
2
H
2
_
1
H
T
1
_
1
= R
1
1
R
1
1
H
1
_
P
1
+H
T
1
R
1
1
H
1
+H
T
2
R
1
2
H
2
_
1
H
T
1
R
1
1
.
The two gain matrices may now be computed from
[ W
1
W
2
] = P
_
H
T
1
H
T
2
_ _
S
11
S
12
S
T
12
S
22
_
1
.
BAE SYSTEMS Proprietary Information, ASCTA use only 80
Substituting in from Equations 156 and 157 we have for the rst gain
W
1
= PH
T
1
1
PH
T
2
_
H
2
PH
T
2
+R
2
_
1
H
2
PH
T
1
1
=
_
PPH
T
2
_
H
2
PH
T
2
+R
2
_
1
H
2
P
_
H
T
1
1
=
_
P
1
+H
T
2
R
1
2
H
2
_
1
H
T
1
1
=
_
P
1
+H
T
2
R
1
2
H
2
_
1
_
1 H
T
1
R
1
1
H
1
_
P
1
+H
T
1
R
1
1
H
1
+H
T
2
R
1
2
H
2
_
1
_
H
T
1
R
1
1
=
_
P
1
+H
T
1
R
1
1
H
1
+H
T
2
R
1
2
H
2
_
1
H
T
1
R
1
1
,
and similarly for the second gain
W
2
=
_
P
1
+H
T
1
R
1
1
H
1
+H
T
2
R
1
2
H
2
_
1
H
T
2
R
1
2
To derive a set of explicit equations for multisensor estimation problems, we could
begin by employing the matrix inversion lemma to invert the innovation matrix for the
twosensor case given in Equation 156, and then proceed to simplify the equation for the
groupsensor gain matrix. However, it is easier in the rst instance to write the gain and
update equations for the groupsensor system in inverse covariance form.
Rewriting the weights associated with the prediction and observation.
I W(k)H(k) = [P(k [ k 1) W(k)H(k)P(k [ k 1)] P
1
(k [ k 1)
= [P(k [ k 1) W(k)S(k) (S
1
(k)H(k)P(k [ k 1))]
P
1
(k [ k 1)
=
_
P(k [ k 1) W(k)S(k)W
T
(k)
_
P
1
(k [ k 1)
= P(k [ k)P
1
(k [ k 1).
(158)
Similarly,
W(k) = P(k [ k 1)H
T
(k)
_
H(k)P(k [ k 1)H
T
(k) +R(k)
_
1
W(k)
_
H(k)P(k [ k 1)H
T
(k) +R(k)
_
= P(k [ k 1)H
T
(k)
BAE SYSTEMS Proprietary Information, ASCTA use only 81
W(k)R(k) = [1 W(k)H(k)] P(k [ k 1)H
T
(k)
so
W(k) = P(k [ k)H
T
(k)R
1
(k). (159)
Substituting Equations 158 and 159 into Equation 90 gives the state update equation as
x(k [ k) = P(k [ k)
_
P
1
(k [ k 1) x(k [ k 1) +H
T
(k)R
1
(k)z(k)
_
. (160)
From Equations 91, 92 and 158 we have
P(k [ k) = [I W(k)H(k)] P(k [ k 1)[I W(k)H(k)]
T
+W(k)R(k)W
T
(k). (161)
Substituting in Equations 158 and 159 gives
P(k [ k) = [P(k [ k)P
1
(k [ k 1)] P(k [ k 1)[P(k [ k)P
1
(k [ k 1)]
T
+
_
P(k [ k)H
T
(k)R
1
(k)
_
R(k)
_
P(k [ k)H
T
(k)R
1
(k)
_
T
.
(162)
Pre and post multiplying by P
1
(k [ k) then simplifying gives the covariance update
equation as
P(k [ k) =
_
P
1
(k [ k 1) +H
T
(k)R
1
(k)H(k)
_
1
. (163)
Thus, in the inversecovariance lter
9
, the state covariance is rst obtained from Equation
163 and then the state itself is found from Equation 160.
Consider now S sensors, each observing a common state according to
z
s
(k) = H
s
(k)x(k) +v
s
(k), s = 1, , S, (164)
where the noise v
s
(k) is assumed to be white and uncorrelated in both time and between
sensors;
Ev
s
(k) = 0,
Ev
s
(i)v
p
(j) =
ij
sp
R
i
(k) s, p = 1, , S; i, j = 1, .
(165)
A composite observation vector z(k) comprising a stacked vector of observations from the
S sensors may be constructed in the form
z(k) =
_
_
z
1
(k)
z
T
S
(k)
_
_ , (166)
a composite observation matrix H(k) comprising the stacked matrix of individual sensor
observation matrices
H(k) =
_
_
H
1
(k)
H
S
(k)
_
_ , (167)
9
Strictly, the inverse covariance lter is dened using the inverse of Equation 163 [37].
BAE SYSTEMS Proprietary Information, ASCTA use only 82
and a composite noise vector v(k) comprising a stacked vector of noise vectors from each
of the sensors
v(k) =
_
_
v
1
(k)
v
T
S
(k)
_
_. (168)
From Equation 165, the covariance in this composite noise vector is a block diagonal
matrix
R(k) = Ev(k)v
T
(k) = blockdiag (R
1
(k), , R
S
(k)) . (169)
With these denitions, we have
H
T
(k)R
1
(k)z(k) = [ H
T
1
(k) H
T
2
(k) H
T
S
(k) ]
_
R
1
1
(k) 0 0
0 R
1
2
(k) 0
.
.
.
.
.
.
.
.
.
.
.
.
0 0 R
1
S
(k)
_
_
_
_
z
1
(k)
z
2
(k)
.
.
.
z
S
(k)
_
_
=
S
i=1
H
T
i
(k)R
1
i
(k)z
i
(k),
(170)
and
H
T
(k)R
1
(k)H(k) = [ H
T
1
(k) H
T
2
(k) H
T
S
(k) ]
_
R
1
1
(k) 0 0
0 R
1
1
(k) 0
.
.
.
.
.
.
.
.
.
.
.
.
0 0 R
1
S
(k)
_
_
_
_
H
1
(k)
H
2
(k)
.
.
.
H
S
(k)
_
_
=
S
i=1
H
T
i
(k)R
1
i
(k)H
i
(k).
(171)
Substituting Equation 170 into Equation 160, the state update equation becomes
x(k [ k) = P(k [ k)
_
P
1
(k [ k 1) x(k [ k 1) +
S
i=1
H
T
i
(k)R
1
i
(k)z
i
(k)
_
. (172)
Substituting Equation 171 into Equation 163, the statecovariance update equation be
comes
P(k [ k) =
_
P
1
(k [ k 1) +
S
i=1
H
T
i
(k)R
1
i
(k)H
i
(k)
_
1
. (173)
The multisensor inversecovariance lter thus consists of a conventional state and state
covariance prediction stage given by Equations 81 and 82, followed by a state and state
covariance update from Equations 172 and 173.
BAE SYSTEMS Proprietary Information, ASCTA use only 83
Example 24
Consider again the multisensor tracking problem of Example 21 with three sensors:
z
1
= [ 1 0 ]
_
x(k)
x(k)
_
+w
1
(k), Ew
2
1
(k) =
2
r
1
,
z
2
= [ 0 1 ]
_
x(k)
x(k)
_
+w
2
(k), Ew
2
2
(k) =
2
r
2
,
and
z
3
= [ 1 0 ]
_
x(k)
x(k)
_
+w
3
(k), Ew
2
3
(k) =
2
r
3
.
Thus the group sensor model is given by
H =
_
_
H
1
H
2
H
3
_
_ =
_
_
1 0
0 1
1 0
_
_,
R =
_
_
R
1
0 0
0 R
2
0
0 0 R
3
_
_ =
_
2
r
1
0 0
0
2
r
2
0
0 0
2
r
3
_
_.
It follows that
H
T
R
1
z =
_
1 0 1
0 1 0
_
_
_
1
2
r
1
0 0
0
1
2
r
2
0
0 0
1
2
r
3
_
_
_
_
z
1
z
2
z
3
_
_ =
_
z
1
2
r
1
+
z
3
2
r
3
z
2
2
r
2
_
=
3
i=1
H
T
i
R
1
i
z
i
,
and
H
T
R
1
H =
_
1 0 1
0 1 0
_
_
_
1
2
r
1
0 0
0
1
2
r
2
0
0 0
1
2
r
3
_
_
_
_
1 0
0 1
1 0
_
_
=
_
_
1
2
r
1
+
1
2
r
3
0
0
1
2
r
2
_
_
=
3
i=1
H
T
i
R
1
i
H
i
,
The advantages of the informationlter form over the groupsensor and asynchronous
approaches to the multisensor estimation problem derive directly from the formulation of
the update equations. Regardless of the number of sensors employed, the largest matrix
inversion required is of dimension the state vector. The addition of new sensors simply
requires that the new terms H
T
i
(k)R
1
i
(k)z
i
(k) and H
T
i
(k)R
1
i
(k)H
i
(k) be constructed
BAE SYSTEMS Proprietary Information, ASCTA use only 84
and added together. Thus the complexity of the update algorithm grows only linearly
with the number of sensors employed. In addition, the update stage can take place in one
single step.
The advantages of the inversecovariance estimator only become obvious when sig
nicant numbers of sensors are employed. It is clear from Equation 173 that both the
prediction covariance and the updated inverse covariance matrix must be inverted in each
cycle of the lter, and so the inversecovariance lter obtains an advantage only when the
dimension of the composite observation vector is approximately two times the dimension
of the common state vector. As we shall see later, it is possible to write the predic
tion stage directly in terms of the inverse covariance, although this does not signicantly
reduce the amount of computation required.
3.2.5 TracktoTrack Fusion
Sensors
z
1
(k) x
1
(kk), P
1
(kk)
^
Local Tracker
z
n
(k) x
n
(kk), P
n
(kk)
^
Local Tracker
z
2
(k) x
2
(kk), P
2
(kk)
^
Local Tracker
TracktoTrack
Fusion
Algorithm
x
T
(kk), P
T
(kk)
^
Figure 24: A typical tracktotrack fusion architecture in which local tracks are generated
at local sensor sites and then communicated to a central fusion centre where a global
track le is assimilated.
Tracktotrack fusion encompasses algorithms which combine estimates from sensor
sites. This is distinct from algorithms that combine observations from dierent sensors;
the former is often called track fusion, the latter scan fusion. In tracktotrack fusion
algorithms, local sensor sites generate local track estimates using a local Kalman lter.
These tracks are then communicated to a central fusion site where they are combined to
generate a global track estimate. A typical tracktotrack fusion architecture is shown
in Figure 24 In some congurations, the global estimate is then communicated back to
the local sensor sites (called a feedback conguration). Tracktotrack fusion algorithms
have a number of potential advantages over scan fusion methods:
1. Local track information is made available for use locally at each sensor sites.
BAE SYSTEMS Proprietary Information, ASCTA use only 85
2. Track information may be communicated at a lower, more compact, rate to the
central site for fusion.
However tracktotrack fusion algorithms also add additional complexity to a system.
Detailed expositions of the tracktotrack fusion method can be found in [2, 11].
The tracktotrack fusion method begins by assuming a common underlying model of
the state being observed in the standard form
x(k) = F(k)x(k 1) +B(k)u(k) +G(k)v(k). (174)
The state is assumed to be observed by a number of sensors each with dierent observation
models in the form
z
i
(k) = H
i
(k)x(k) +w
i
(k), i = 1, 2, (175)
but where observation noises are assumed independent.
Ew
i
(k)w
j
(k) =
ij
R
i
(k). (176)
A local track is formed by each sensor node, on the basis of only local observations, using
the normal Kalman lter algorithm as
x
i
(k [ k) = x
i
(k [ k 1) +W
i
(k) [z
i
(k) H
i
(k) x
i
(k [ k 1)] , (177)
and
P
i
(k [ k) = P
i
(k [ k 1) W
i
(k)S
i
(k)W
T
i
(k), (178)
where
W
i
(k) = P
i
(k [ k 1)H
T
i
(k)S
1
i
(k), (179)
and
S
i
(k) =
_
H
i
(k)P
i
(k [ k 1)H
T
i
(k) +R
i
(k)
_
. (180)
Local state predictions are generated from a common state model
x
i
(k [ k 1) = F(k) x
i
(k 1 [ k 1) +B(k)u(k) (181)
and
P
i
(k [ k 1) = F(k)P
i
(k 1 [ k 1)F
T
(k) +G(k)Q(k)G
T
(k) (182)
A straightforward track fusion algorithm is simply to take the variance weighted
average of tracks as follows:
x
T
(k [ k) = P
T
(k [ k)
N
i=1
P
1
i
(k [ k) x
i
(k [ k) (183)
P
T
(k [ k) =
_
N
i=1
P
1
i
(k [ k)
_
1
. (184)
This is a commonly practically used tracktotrack fusion method.
BAE SYSTEMS Proprietary Information, ASCTA use only 86
However, a central issue in tracktotrack fusion is that any two tracks x
i
(k [ k) and
x
j
(k [ k) are correlated because they have a common prediction error resulting from a
common process model. This correlation is intrinsic to the problem; it is only because the
states have this process model in common that there is a reason to fuse the two. Thus it
is not possible in general to compute a fused track x
T
(k [ k) from a simple linear weighted
sum of local tracks.
To address the general tracktotrack fusion problem, it is necessary to nd an expres
sion for the correlation between two tracks. When this is known, it is possible to fuse the
two estimates. To begin, we compute an expression for the track estimate and prediction
errors as
x
i
(k [ k) = x(k) x
i
(k [ k)
= x(k) x
i
(k [ k 1) W
i
(k) [z
i
(k) H
i
(k) x
i
(k [ k 1)]
= x(k) x
i
(k [ k 1) W
i
(k) [H
i
(k)x(k) +w
i
(k) H
i
(k) x
i
(k [ k 1)]
= [1 W
i
(k)H
i
(k)] x
i
(k [ k 1) W
i
(k)w
i
(k) (185)
and
x
i
(k [ k 1) = x(k) x
i
(k [ k 1)
= F(k)x(k 1) F(k) x
i
(k 1 [ k 1) +G(k)v(k)
= F(k) x
i
(k 1 [ k 1) +G(k)v(k) (186)
Squaring and taking expectations of Equation 186 yields and expression for the predicted
tracktotrack crosscorrelation
P
ij
(k [ k 1) = E x
i
(k [ k 1) x
T
j
(k [ k 1) [ Z
k1
= E[F(k) x
i
(k 1 [ k 1) +G(k)v(k)]
[F(k) x
i
(k 1 [ k 1) +G(k)v(k)]
T
= F(k)E x
i
(k 1 [ k 1) x
T
j
(k 1 [ k 1) [ Z
k1
F
T
(k)
+G(k)Ev(k)v
T
(k) G
T
(k)
= F(k)P
ij
(k 1 [ k 1)F
T
(k) +G(k)Q(k)G
T
(k). (187)
Squaring and taking expectations of Equation 185 yields an expression for the estimate
tracktotrack crosscorrelation
P
ij
(k [ k) = E x
i
(k [ k) x
T
j
(k [ k)
= E[(1 W
i
(k)H
i
(k)) x
i
(k [ k 1) W
i
(k)w
i
(k)]
[(1 W
j
(k)H
j
(k)) x
j
(k [ k 1) W
j
(k)w
j
(k)] [ Z
k
= [1 W
i
(k)H
i
(k)] E x
i
(k [ k 1) x
T
j
(k [ k 1) [1 W
j
(k)H
j
(k)]
T
+W
i
(k)Ew
i
(k)w
T
j
(k) W
T
j
(k)
= [1 W
i
(k)H
i
(k)] P
ij
(k [ k 1) [1 W
j
(k)H
j
(k)]
T
(188)
BAE SYSTEMS Proprietary Information, ASCTA use only 87
where we have used the fact that Ew
i
(k)w
T
j
(k) = 0. Together, Equations 187 and 188
provide a recursive relationship for computing the crosscorrelation P
ij
(k [ k) between the
two track estimates x
i
(k [ k) and x
j
(k [ k).
Fusing together two tracks is essentially the same as adding observation information
except that the data are correlated. Recall
x(k [ k) = x(k [ k 1) +P
xz
P
1
zz
[z(k) z(k [ k 1)] (189)
and
P(k [ k) = P(k [ k 1) P
xz
P
1
zz
P
T
xz
(190)
so
x
T
(k [ k) = x
i
(k [ k) +P
ij
(k [ k)P
1
i+j
(k [ k) [ x
j
(k [ k) x
i
(k [ k)] (191)
and
P
T
(k [ k) = P
i
(k [ k) P
ij
(k [ k)P
1
i+j
(k [ k)P
T
ij
(k [ k) (192)
identify
P
ij
(k [ k) = P
i
(k [ k) P
ij
(k [ k) (193)
and
P
i+j
(k [ k) = P
i
(k [ k) +P
j
(k [ k) P
ij
(k [ k) P
T
ij
(k [ k) (194)
Substituting Equations 193 and 194 into Equation 191 gives
x
T
(k [ k) = x
i
(k [ k) + [P
i
(k [ k) P
ij
(k [ k)]
_
P
i
(k [ k) +P
j
(k [ k) P
ij
(k [ k) P
T
ij
(k [ k)
_
1
[ x
j
(k [ k) x
i
(k [ k)] (195)
as the combined track estimate, and
P
T
(k [ k) = P
i
(k [ k) [P
i
(k [ k) P
ij
(k [ k)]
_
P
i
(k [ k) +P
j
(k [ k) P
ij
(k [ k) P
T
ij
(k [ k)
_
1
[P
i
(k [ k) P
ij
(k [ k)]
T
(196)
Equations 195 and 196 are in the form of predictorcorrector equations. As written,
x
i
(k [ k) is the predicted track, and x
j
(k [ k) x
i
(k [ k) is the correction term, weighted
by a gain proportional to the corresponding covariances. The equations are symmetric so
that the role of i and j are interchangeable.
There are a number of extensions to the basic tracktotrack fusion algorithm. Notable
is the use of equivalent measurements described in [11]. However, these methods are
more appropriately described in context of the information lter.
BAE SYSTEMS Proprietary Information, ASCTA use only 88
4 Decentralised Data Fusion Systems
The section addresses the development of algorithms for decentralised data fusion archi
tectures.
The nature of data fusion is that there are a number of sensors physically distributed
around an environment. In a centralised data fusion system, raw sensor information is
communicated back to a central processor where the information is combined to produce
a single fused picture of the environment. In a distributed data fusion system, each
sensor has its own local processor which can generally extract useful information from
the raw sensor data prior to communication. This has the advantage that less information
is normally communicated, the computational load on the central processor is reduced
and the sensors themselves can be constructed in a reasonably modular manner. The
degree to which local processing occurs at a sensor site varies substantially from simple
validation and data compression up to the full construction of tracks or local interpretation
of information.
While for many systems a centralised approach to data fusion is adequate, the increas
ing sophistication, functional requirements, complexity and size of data fusion systems,
coupled with the ever reducing cost of computing power argues more and more toward
some form of distributed processing. The central issue in designing distributed data fusion
systems is the development of appropriate algorithms which can operate at a number of
distributed sites in a consistent manner. This is the focus of this section.
This section begins with a general discussion of data fusion architectures and the chal
lenges posed in developing distributed data fusion algorithms. The information lter, and
more generally the loglikelihood implementations of Bayes theorem are then developed
and it is shown how these can readily be mapped to many distributed and decentralised
data fusion systems.
4.1 Data Fusion Architectures
Distributed data fusion systems may take many forms. At the simplest level, sensors could
communicate information directly to a central processor where it is combined. Little or
no local processing of information need take place and the relative advantage of having
many sources of information is sacriced to having complete centralised control over the
processing and interpretation of this information. As more processing occurs locally, so
computational and communication burden can be removed from the fusion center, but at
the cost of reduced direct control of lowlevel sensor information.
Increasing intelligence of local sensor nodes naturally results in a hierarchical structure
for the fusion architecture. This has the advantage of imposing some order on the fusion
process, but the disadvantage of placing a specic and often rigid structure on the fusion
system.
Other distributed architectures consider sensor nodes with signicant local ability to
generate tracks and engage in fusion tasks. Such architectures include Blackboard and
agent based systems.
BAE SYSTEMS Proprietary Information, ASCTA use only 89
Fully decentralised architectures have no central processor and no common communi
cation system. In such systems, nodes can operate in a fully autonomous manner, only
coordinating through the anonymous communication information.
The following sections consider these architectures and their associated data fusion
algorithms.
4.1.1 Hierarchical Data Fusion Architectures
In a hierarchical structure, the lowest level processing elements transmit information
upwards, through successive levels, where the information is combined and rened, until
at the top level some global view of the state of the system is made available. Such
hierarchical structures are common in many organisations and have many wellknown
advantages over fully centralised systems; particularly in reducing the load on a centralised
processor while maintaining strict control over subprocessor operations.
z
1
(k)
z
2
(k)
z
N
(k)
Sensors
Tracking
System 2
Tracking
System 1
Tracking
System N
x
1
(k)
^
x
2
(k)
^
x
N
(k)
^
Tracks
Track Fusion
Algorithms
Combined
Track
Figure 25: Single Level Hierarchical Multiple Sensor Tracking System
The hierarchical approach to systems design has been employed in a number of data
fusion systems and has resulted in a variety of useful algorithms for combining information
at dierent levels of a hierarchical structure. General hierarchical Bayesian algorithms are
based on the independent likelihood pool architectures shown in Figures 8 and 9, or on the
loglikelihood opinion pools shown in Figures 12 and 13. Here the focus is on hierarchical
estimation and tracking algorithms (See Figures 25 and 26).
First it is assumed that all sensors are observing a common state or track x(k). Ob
servations are made at local sites of this common state according to a local observation
equation in the form
z
i
(k) = H
i
(k)x(k) +w
i
(k), i = 1, , S (197)
In principle, each site may then operate a conventional Kalman lter or state estimator
to provide local state estimates based only on local observations in the form
x
i
(k [ k) = x
i
(k [ k 1) +W
i
(k) [z
i
(k) H
i
(k) x
i
(k [ k 1)] , i = 1, , S. (198)
BAE SYSTEMS Proprietary Information, ASCTA use only 90
Sitelevel
Track Fusion
Group
Picture
Tracks
z
1
(k)
z
n
(k)
Tracking
System n
Tracking
System 1
x
1
(k)
^
x
n
(k)
^
z
1
(k)
z
n
(k)
Tracking
System n
Tracking
System 1
x
1
(k)
^
x
n
(k)
^
z
1
(k)
z
n
(k)
Tracking
System n
Tracking
System 1
x
1
(k)
^
x
n
(k)
^
Combined
Picture
Tracks
Group
Picture
Tracks
Figure 26: Multiple Level Hierarchical Multiple Sensor Tracking System
BAE SYSTEMS Proprietary Information, ASCTA use only 91
These local estimates x
i
(k [ k) may then be passed further up the hierarchy to a central
or intermediate processor which combines or fuses tracks to form a global estimate based
on all observations in the form
x(k [ k) =
S
i=1
i
(k) x
i
(k [ k), (199)
where
i
(k) are site weighting matrices.
The essential problem, as described in Section 3, is that each sensor site must be
observing a common true state and so the local process models are related through some
common global model in the form
x
i
(k) = F
i
(k)x(k) +B
i
(k)u(k) +G
i
(k)v(k), i = 1, , S.
This means that the predictions made at local sites are correlated and so the updated local
estimates in Equation 198 must also be correlated despite the fact that the observations
made at each site are dierent. Consequently, the local estimates generated at each site
cannot be combined in the independent fashion implied by Equation 199.
The correct method of dealing with this problem is to explicitly account for these
correlations in the calculation of the site weighting matrices
i
(k). In particular, the
tracktotrack fusion algorithms described in Section 3.2.5 in the form of Equations 195
and 196 are appropriate to this problem. These algorithms require that the correlations
between all sites be explicitly computed in addition to the covariance associated with each
local state estimate.
There have been a number of papers on hierarchical estimation systems. The paper by
Hashemipour, Roy and Laub [28] is notable in employing, indirectly, the information form
of the Kalman lter to derive a hierarchical estimation algorithm. The earlier paper by
Speyer [54] has a similar formulation, and although it is concerned with distributed linear
quadratic Gaussian (LQG) control problems, also specically deals with communication or
transmission requirements. Other large scale systems in control exhibit similar properties
[52]. In addition, the tracktotrack fusion techniques described by BarShalom [16, 2]
serve as the basis for many derived architectures.
A hierarchical approach to the design of data fusion systems also comes with a number
of inherent disadvantages. The ultimate reliance on some central processor or controlling
level within the hierarchy means that reliability and exibility are often compromised.
Failure of this central unit leads to failure of the whole system, changes in the system often
mean changes in both the central unit and in all related subunits. Further, the burden
placed on the central unit in terms of combining information can often still be prohibitive
and lead to an inability of the design methodology to be extended to incorporate an
increasing number of sources of information. Finally, the inability of information sources
to communicate, other than through some higher level in the hierarchy, eliminates the
possibility of any synergy being developed between two or more complimentary sources
of information and restricts the system designer to rigid predetermined combinations of
information. The limitations imposed by a strict hierarchy have been widely recognised
both in human information processing systems as well as in computerbased systems.
BAE SYSTEMS Proprietary Information, ASCTA use only 92
4.1.2 Distributed Data Fusion Architectures
The move to more distributed, autonomous, organisations is clear in many information
processing systems. This is most often motivated by two main considerations; the desire
to make the system more modular and exible, and a recognition that a centralised
or hierarchical structure imposes unacceptable overheads on communication and central
computation. The migration to distributed system organisations is most apparent in
Articial Intelligence (AI) application areas, where distributed AI has become a research
area in its own right. Many of the most interesting distributed processing organisations
have originated in this area.
Track Fusion
Algorithms
Blackboard
Medium
z
1
(k)
Tracking
System 1
z
2
(k)
Tracking
System 2
Track Identity
Knowledge Base
Situation
Knowledge Base
Remote Sensor
Knowledge Base
Figure 27: Blackboard Architecture in Data Fusion
Notable is the Blackboard architecture (see Figure 27), originally developed in the
Hearsay speech understanding programme, but now widely employed in many areas of AI
and data fusion research. A Blackboard architecture consists of a number of independent
autonomous agents. Each agent represents a source of expert knowledge or specialised
information processing capability. Agents exchange information through a common com
munication facility or shared memory resource. This resource is called a blackboard. The
blackboard is designed to closely replicate its physical analogue. Each agent is able to
write information or local knowledge to this resource. Every agent in the system is able
to read from this resource, in an unrestricted manner, any information which it considers
useful in its current task. In principle, every agent can be made modular and new agents
may be added to the system when needed without changing the underlying architecture or
operation of the system as a whole. The exibility of this approach to system organisation
has made the Blackboard architecture popular in a range of application domains [45]. In
data fusion, the Blackboard approach has been most widely used for knowledgebased
BAE SYSTEMS Proprietary Information, ASCTA use only 93
data fusion systems in data interpretation and situation assessment (see [26] and [27] for
example). However, the structured nature of tracking and identication problems does
not lend itself to this anarchic organisational form.
The Blackboard architecture has a number of basic problems. All of these stem from
the use of a common communication or memory resource. The core problem is that a
central resource naturally entails the need for some type of central control in which a single
decision maker is used to sequence and organise the reading and writing of information
from the shared resource. Practically, with such a control mechanism, a blackboard
architecture becomes no more than a one level hierarchy with consequent lack of exibility
and with the inherent limitations imposed by the use of a central resource.
4.1.3 Decentralised Data Fusion Architectures
A decentralized data fusion system consists of a network of sensor nodes, each with its
own processing facility, which together do not require any central fusion or central com
munication facility. In such a system, fusion occurs locally at each node on the basis of
local observations and the information communicated from neighbouring nodes. At no
point is there a common place where fusion or global decisions are made.
A decentralised data fusion system is characterised by three constraints:
1. There is no single central fusion center; no one node should be central to the suc
cessful operation of the network.
2. There is no common communication facility; nodes cannot broadcast results and
communication must be kept on a strictly nodetonode basis.
3. Sensor nodes do not have any global knowledge of sensor network topology; nodes
should only know about connections in their own neighbourhood.
Figures 28 and 29 and 30 show three possible realisations of a decentralised data fusion
system. The key point is that all these systems have no central fusion center (unlike
the decentralised systems often described in the literature which are actually typically
distributed or hierarchical).
The constraints imposed provide a number of important characteristics for decen
tralised data fusion systems:
Eliminating the central fusion center and any common communication facility en
sures that the system is scalable as there are no limits imposed by centralized
computational bottlenecks or lack of communication bandwidth.
Ensuring that no node is central and that no global knowledge of the network
topology is required for fusion means that the system can be made survivable to
the online loss (or addition) of sensing nodes and to dynamic changes in the network
structure.
BAE SYSTEMS Proprietary Information, ASCTA use only 94
As all fusion processes must take place locally at each sensor site and no global
knowledge of the network is required a priori, nodes can be constructed and pro
grammed in a modular fashion.
These characteristics give decentralised systems a major advantage over more traditional
sensing architectures, particularly in defense applications.
Sensor Node
Sensor
Fusion Processor
Communications Medium
Figure 28: A decentralised data fusion system implemented with a pointtopoint com
munication architecture.
A decentralized organization diers from a distributed processing system in having
no central processing or communication facilities. Each sensor node in a decentralized
organization is entirely self contained and can operate completely independently of any
other component in the system. Communication between nodes is strictly onetoone
and requires no remote knowledge of node capability. Throughout this section we distin
guish between decentralized organizations that have no common resources, and distributed
organizations where some residual centralized facility is maintained.
4.2 Decentralised Estimation
Decentralised data fusion is based on the idea of using formal information measures as
the means of quantifying, communicating and assimilating sensory data. In decentralised
BAE SYSTEMS Proprietary Information, ASCTA use only 95
Communications Medium
Sensor Node
Sensor
Fusion Processor
Figure 29: A decentralised data fusion system implemented with a broadcast, fully con
nected, communication architecture. Technically, a common communication facility vio
lates decentralised data fusion constraints. However a broadcast medium is often a good
model of real communication networks.
Sensor Payloads
Internal Communciation
External Communication
Figure 30: A decentralised data fusion system implemented with a hybrid, broadcast and
pointtopoint, communication architecture.
BAE SYSTEMS Proprietary Information, ASCTA use only 96
estimation of continuous valued states, this is implemented in the form of an information
lter. In this section, the full form of the information lter is derived. It is then demon
strated how the lter may be decentralised amongst a number of sensing nodes. Later
sections then describe how to deal with issues of communication and data association in
decentralised sensing
4.2.1 The Information Filter
Conventional Kalman lters deal with the estimation of states x(i), and yield estimates
x(i [ j) together with a corresponding estimate variance P(i [ j). The information lter
deals instead with the information state vector y(i [ j) and information matrix Y(i [ j)
dened as
y(i [ j) = P
1
(i [ j) x(i [ j), Y(i [ j) = P
1
(i [ j). (200)
These information quantities have an interpretation related to the underlying probability
distributions associated with the estimation problem. The information matrix in par
ticular is closely associated with the Fisher information measures introduced in Section
2.
A set of recursion equations for the information state and information matrix can be
derived directly from the equations for the Kalman lter. The resulting information lter
is mathematically identical to the conventional Kalman lter.
Recall the update stage for the Kalman lter:
x(k [ k) = (1 W(k)H(k)) x(k [ k 1) +W(k)z(k) (201)
P(k [ k) = (1 W(k)H(k))P(k [ k 1)(1 W(k)H(k))
T
+W(k)R(k)W
T
(k) (202)
Now, from Equations 158 and 159, we have
1 W(k)H(k) = P(k [ k)P
1
(k [ k 1), (203)
and
W(k) = P(k [ k)H
T
(k)R
1
(k). (204)
Substituting Equations 203 and 204 into Equation 201 gives
x(k [ k) = P(k [ k)P
1
(k [ k 1) x(k [ k 1) +P(k [ k)H
T
(k)R
1
(k)z(k).
Premultiplying through by P
1
(k [ k) gives the update equation for the informationstate
vector as
P
1
(k [ k) x(k [ k) = P
1
(k [ k 1) x(k [ k 1) +H
T
(k)R
1
(k)z(k). (205)
Dening
i(k)
= H
T
(k)R
1
(k)z(k) (206)
BAE SYSTEMS Proprietary Information, ASCTA use only 97
as the informationstate contribution from an observation z(k), and with the denitions
in Equation 200, Equation 205 becomes
y(k [ k) = y(k [ k 1) +i(k). (207)
A similar expression can be obtained for the covariance update of Equation 202. Substi
tuting Equations 203 and 204 into Equation 202 and rearranging gives
P
1
(k [ k) = P
1
(k [ k 1) +H
T
(k)R
1
(k)H(k). (208)
Dening
I(k)
= H
T
(k)R
1
(k)H(k) (209)
as the information matrix associated with the observation, and with the denitions in
Equation 200, Equation 208 becomes the information matrix update equation
Y(k [ k) = Y(k [ k 1) +I(k). (210)
A comparison of the Kalman lter update stage (Equations 201 and 202) with the in
formation lter update stage (Equations 207 and 210) highlights the simplicity of the
information lter update over the Kalman lter update. Indeed, in information form, the
update stage is a straight addition of information from a prediction and from an observa
tion. It is this simplicity which gives the information lter its advantage in multisensor
estimation problems.
The simplicity of the update stage of the information lter comes at the cost of in
creased complexity in the prediction stage. Recall the covariance prediction equation
P(k [ k 1) = F(k)P(k 1 [ k 1)F
T
(k) +G(k)Q(k)G
T
(k) (211)
To derive the prediction stage for the information lter, the following version of the matrix
inversion lemma is noted [37]
_
A+B
T
C
_
1
= A
1
A
1
B
T
_
1 +CA
1
B
T
_
1
CA
1
.
Identifying
A = F(k)P(k 1 [ k 1)F
T
(k), B
T
= G(k)Q(k), C = G
T
(k),
the inverse of Equation 211 becomes
P
1
(k [ k 1) = M(k) M(k)G(k)
_
G
T
(k)M(k)G(k) +Q
1
(k)
_
1
G
T
(k)M(k) (212)
where
M(k) = F
T
(k)P
1
(k 1 [ k 1)F
1
(k) (213)
when Q(k) is non singular. Noting the denition of the state transition matrix
F(k)
= (t
k
, t
k1
)
BAE SYSTEMS Proprietary Information, ASCTA use only 98
implies F
1
(k) always exists and indeed
F
1
(k) = (t
k1
, t
k
)
is simply the state transition matrix dened backwards from a time t
k
to t
k1
. Now,
dening
(k)
=
_
G
T
(k)M(k)G(k) +Q
1
(k)
_
, (214)
and
(k)
= M(k)G(k)
_
G
T
(k)M(k)G(k) +Q
1
(k)
_
1
= M(k)G(k)
1
(k), (215)
the information matrix prediction equation becomes
Y(k [ k 1) = P
1
(k [ k 1) = M(k) (k)(k)
T
(k). (216)
A number of alternate expressions for the information prediction stage can also be derived:
Y(k [ k 1) =
_
1 (k)G
T
(k)
_
M(k) (217)
and
Y(k [ k 1) =
_
1 (k)G
T
_
M(k)
_
1 (k)G
T
_
T
+(k)Q
1
(k)
T
(k). (218)
The informationstate prediction equations may also be obtained as
y(k [ k 1) =
_
1 (k)G
T
(k)
_
F
T
(k)
[ y(k 1 [ k 1) +Y(k 1 [ k 1)F
1
(k)B(k)u(k)]
=
_
1 (k)G
T
(k)
_ _
F
T
(k) y(k 1 [ k 1) +M(k)B(k)u(k)
_
=
_
1 (k)G
T
(k)
_
F
T
(k) y(k 1 [ k 1)
+Y(k [ k 1)B(k)u(k). (219)
It should be noted that the complexity of the inversion of (k) is only of order the
dimension of the driving noise (often scalar). Further Q(k) is almost never singular, and
indeed if it were, singularity can be eliminated by appropriate denition of G(k). The
special case in which Q(k) = 0 yields prediction equations in the form:
Y(k [ k 1) = M(k), y(k [ k 1) = F
T
(k) y(k 1 [ k 1)+M(k)B(k)u(k) (220)
The information lter is now summarised
Prediction:
y(k [ k 1) =
_
1 (k)G
T
(k)
_
F
T
(k) y(k 1 [ k 1) +Y(k [ k 1)B(k)u(k) (221)
Y(k [ k 1) = M(k) (k)(k)
T
(k) (222)
BAE SYSTEMS Proprietary Information, ASCTA use only 99
where
M(k) = F
T
(k)Y(k 1 [ k 1)F
1
(k), (223)
(k) = M(k)G(k)
1
(k), (224)
and
(k) =
_
G
T
(k)M(k)G(k) +Q
1
(k)
_
. (225)
Estimate:
y(k [ k) = y(k [ k 1) +i(k) (226)
Y(k [ k) = Y(k [ k 1) +I(k). (227)
where
i(k) = H
T
(k)R
1
(k)z(k), I(k) = H
T
(k)R
1
(k)H(k) (228)
Given the interpretation of the information matrix Y(i [ j) as the Fisher information, the
update Equation 227 is simply seen to add the information contributed by the observa
tion. Conversely, the prediction Equation 222 subtracts information caused by process
uncertainties.
Example 25
Consider again Example 16 of constant velocity particle motion:
_
x(t)
x(t)
_
=
_
0 1
0 0
_ _
x(t)
x(t)
_
+
_
0
1
_
v(t).
where v(t) is a zero mean white noise process with Ev
2
(t) = q. The state transition
matrix and its inverse over a time interval tare given by
F(t) =
_
1 t
0 1
_
, F
1
(t) =
_
1 t
0 1
_
and the noise transition matrix by (Equation 69)
G(t) =
_
t
0
_
1
0 1
_ _
0
1
_
d =
_
t
2
/2
t
_
Setting
Y(k 1 [ k 1) =
_
Y
xx
Y
xv
Y
xv
Y
vv
_
,
the propagated information is
M(k) = F
T
(k)Y(k 1 [ k 1)F
1
(k) =
_
Y
xx
Y
xv
Y
xx
t
Y
xv
Y
xx
Y
xx
t
2
2Y
xv
t +Y
vv
_
.
Then
(k) =
_
G
T
(k)M(k)G(k) +Q
1
(k)
_
= t
2
_
Y
xx
t
2
/4 Y
xv
t +Y
vv
_
+q
1
BAE SYSTEMS Proprietary Information, ASCTA use only 100
and
(k) = M(k)G(k)
1
(k)
=
1
Y
xx
t
4
/4 Y
xv
t
3
+Y
vv
t
2
+q
1
_
Y
xv
t Y
xx
t
2
/2
Y
xx
t
3
/2 3Y
xv
t
2
/2 +Y
vv
_
.
Finally, denoting
M(k) =
_
M
xx
M
xv
M
xv
M
vv
_
,
the propagation gain matrix is found as
1 (k)G
T
(k) =
1
t
2
4
M
xx
tM
xv
+M
vv
+
q
1
t
2
_
M
xv
t/2 +M
vv
+q
1
/t
2
M
xx
t/2 +M
xv
M
xv
t
2
/4 +M
vv
t/2 M
xx
+M
xv
t/2 +q
1
/t
2
_
.
Note also the term
F
T
(k) y(k 1 [ k 1) =
_
1 0
t 1
_ _
y
y
_
=
_
y
y ty
_
Assume observations are made of the position of the particle at discrete synchronous
time intervals according to
z(k) = [ 1 0 ]
_
x(k)
x(k)
_
+w(k)
where w(t) is a zero mean white noise process with Ew
2
(t) = r. The information state
is simply given by
i(k) = H
T
(k)R
1
(k)z(k) =
_
1
0
_
r
1
z
x
=
_
z
x
/r
0
_
,
and the information matrix by
I(k) = H
T
(k)R
1
(k)H(k) =
_
1
0
_
r
1
[ 1 0 ] =
_
1/r 0
0 0
_
.
A key problem with this information state vector y(i [ j) is that it has no obvious
metric properties; the dierence between two informationstate vectors does not relate at
all to the dierence in state vectors because of scaling by the information matrix. This
can make the interpretation of the information lter more dicult than the (statebased)
Kalman lter.
BAE SYSTEMS Proprietary Information, ASCTA use only 101
There is a clear duality between the information lter and the conventional Kalman
lter, in which the prediction stage of the information lter is related to the update stage
of the Kalman lter, and the update stage of the information lter to the prediction stage
of the Kalman lter [1]. In particular, identifying the correspondences
(k) W(k), (k) S(k), G
T
(k) H(k)
(k) is seen to take on the role of an information prediction gain matrix, (k) an in
formation innovation matrix, and G
T
(k) an information observation. This duality is
instructive in understanding the relationship of information to state as equivalent repre
sentations. It is also of assistance in implementation of the information ltering equations.
With this interpretation, The informationlter form has the advantage that the up
date Equations 226 and 227 for the estimator are computationally simpler than the cor
responding equations for the Kalman Filter, at the cost of increased complexity in pre
diction. The value of this in decentralized sensing is that estimation occurs locally at
each node, requiring partition of the estimation equations which are simpler in their in
formation form. Prediction, which is more complex in this form, relies on a propagation
coecient which is independent of the observations made and so is again simpler to de
couple and decentralize amongst a network of sensor nodes. This property is exploited in
subsequent sections.
4.2.2 The Information Filter and Bayes Theorem
There is a strong relationship between the information state and the underlying proba
bility distribution. Recall Bayes Theorem
P(x(k) [ Z
k
) =
P(z(k) [ x(k))P(x(k) [ Z
k1
)
P(z(k) [ Z
k1
)
. (229)
If it is assumed that the prior and likelihood are Gaussian as
P(x(k) [ Z
k1
) exp
_
1
2
[x(k) x(k [ k 1)]
T
P
1
(k [ k 1) [x(k) x(k [ k 1)]
_
.
P(z(k) [ x(k)) exp
_
1
2
[z(k) H(k)x(k)]
T
R
1
(k) [z(k) H(k)x(k)]
_
,
then the posterior will also be Gaussian in the form
P(x(k) [ Z
k
) exp
_
1
2
[x(k) x(k [ k)]
T
P
1
(k [ k) [x(k) x(k [ k)]
_
Now, substituting these distributions into Equation 229 and take logs gives
[x(k) x(k [ k)]
T
P
1
(k [ k) [x(k) x(k [ k)] =
[z(k) H(k)x(k)]
T
R
1
(k) [z(k) H(k)x(k)]
+ [x(k) x(k [ k 1)]
T
P
1
(k [ k 1) [x(k) x(k [ k 1)] +C(z(k)) (230)
BAE SYSTEMS Proprietary Information, ASCTA use only 102
where C(z(k)) is independent of x(k). This equation relates the log likelihoods of these
Gaussian distributions and is essentially a quadratic in x(k). Now, dierentiating this
expression once with respect to x(k) and rearranging gives Equation 226. Dierentiating
a second time and rearranging gives
P
1
(k [ k) = H
T
(k)R
1
(k)H(k) +P
1
(k [ k 1),
which is Equation 227. The rst derivative of the loglikelihood is known as the score
function. The second derivative is, of course, the Fisher information Equation 59.
This analysis suggests that the information lter is, fundamentally, a log likelihood
implementation of Bayes Theorem. The rst and second derivatives of the log likelihood
are essentially moment generating functions (the rst derivative is the centre of mass,
and the second, the moment of inertia [46]). The information lter is thus a means for
recursive calculation of sucient statistics (the mean and variance) in the case when the
distributions of interest our Gaussian.
This interpretation of the information lter shows the relationship between the Kalman
lter and the more general probabilistic estimation problem. Indeed, if the distributions
were not Gaussian and indeed possibly discrete, the information lter would reduce to
the general fusion of log likelihoods as described in Section 2.2.6. This can be exploited
in the development of ecient distributed data fusion algorithms for problems which are
not linear or Gaussian.
4.2.3 The Information lter in MultiSensor Estimation
The information lter is reasonably well known in the literature on estimation [1, 37].
However, its use in data fusion has been largely neglected in favour of conventional state
based Kalman ltering methods. The reasons for this appear to be somewhat spurious,
based largely on the incorrect hypothesis that it is cheaper to communicate innovation
information (of dimension the observation vector) than to communicate information state
vectors (of dimension the state vector). The basic problem is that it is generally not possi
ble to extend Equations 201 and 202 in any simple way to deal with multiple observations.
The reason for this, as we have seen, is that although the innovation vectors at dierent
times are uncorrelated (by construction), the innovations generated by dierent sensors
at the same time are correlated, by virtue of the fact that they use a common prediction.
Thus, the innovation covariance matrix S(k) can never be diagonal, and so can not be
simply partitioned and inverted to yield a gain matrix for each individual observation.
Thus, for a set of sensors i = 1, , N, it is not possible to compute the simple sum of
innovations as
x(k [ k) ,= x(k [ k 1) +
N
i=1
W
i
(k) [z
i
(k) H
i
(k) x(k [ k 1)]
in which the individual gains are given by
W
i
(k) = P(k [ k 1)H
T
i
(k)S
1
i
(k).
BAE SYSTEMS Proprietary Information, ASCTA use only 103
However, the information lter provides a direct means of overcoming these problems.
As will be seen, for the same set of sensors, i = 1, , N, and without any additional
assumptions, it is the case that the information contributions from all sensors can simply
be added to obtain an updated estimate in the form
y(k [ k) = y(k [ k 1) +
N
j=1
i
j
(k),
which is algebraically equivalent to a full (all sensor) statebased Kalman lter estimate.
The reason why this can be done with the information lter is that the information
contributions made by the observations are directly related to the underlying likelihood
functions for the states rather than to the state estimates themselves. This can be appreci
ated by considering the interpretation of the information lter directly an implementation
of Bayes theorem in terms of loglikelihood rather than in terms of state. Indeed, making
the usual assumption that the observations made by the various sensors are conditionally
independent given the true state as
P(z
1
(k) , z
N
(k) [ x(k)) =
N
i=1
P(z
i
(k) [ x(k)), (231)
then Bayes Theorem gives (Equation 19)
P(x(k) [ Z
n
(k)) = P(x(k) [ Z
n
(k 1))
n
i=1
P(z
i
(k) [ x(k)).[P(Z
n
(k) [ Z
n
(k 1))]
1
.
(232)
Taking logs of Equation 232 then gives
ln P(x(k) [ Z
n
(k)) = ln P(x(k) [ Z
n
(k 1))
+
n
i=1
ln P(z
i
(k) [ x(k)) ln P(Z
n
(k) [ Z
n
(k 1)). (233)
Given Equation 230, an identication can be made as
n
i=1
ln P(z
i
(k) [ x(k))
n
i=1
i(k),
n
i=1
I(k) . (234)
This demonstrates why, fundamentally, it is possible to add information states and Infor
mation matrices from dierent sensors while it is not possible to add innovations without
accounting for crosscorrelations. For this reason also, the information lter is occasionally
referred to as the likelihood lter.
Computationally, the information lter thus provides a far more natural means of as
similating information than does the conventional Kalman lter and a far simpler method
of dealing with complex multisensor data fusion problems.
BAE SYSTEMS Proprietary Information, ASCTA use only 104
The linear addition of information in the information lter can also be obtained by
direct algebraic manipulation. Consider a system comprising N sensors each taking ob
servations according to
z
i
(k) = H
i
(k)x(k) +v
i
(k) (235)
with
Ev
p
(i)v
T
q
(j) =
ij
pq
R
p
(i). (236)
The observations can be stacked into a composite observation
z(k) =
_
z
T
1
(k), , z
T
N
(k)
_
T
(237)
The observation model can also be stacked into a composite model
H(k) =
_
H
T
1
(k), , H
T
N
(k)
_
T
, (238)
and
v(k) =
_
v
T
1
(k), , v
T
N
(k)
_
T
(239)
to give a composite observation equation in familiar form
z(k) = H(k)x(k) +v(k).
Noting that
Ev(k)v
T
(k) = R(k) = blockdiagR
T
1
(k), , R
T
N
(k) , (240)
a multiple sensor form of Equation 228 can be obtained as
i(k) =
N
i=1
i
i
(k) =
N
i=1
H
T
i
(k)R
1
i
(k)z
i
(k) (241)
and
I(k) =
N
i=1
I
i
(k) =
N
i=1
H
T
i
(k)R
1
i
(k)H
i
(k) (242)
where
i
i
(k)
= H
T
i
(k)R
1
i
(k)z
i
(k) (243)
is the informationstate contribution from observation z
i
(k) and
I
i
(k)
= H
T
i
(k)R
1
i
(k)H
i
(k) (244)
is its associated information matrix.
Equations 241 and 242 show that the total information available to the lter at any
timestep is simply the sum of the information contributions from each of the individual
BAE SYSTEMS Proprietary Information, ASCTA use only 105
sensors. Further, Equations 226 and 227 describing the single sensor information update
are easily extended to multiple sensor updates in a straightforward manner as
y(k [ k) = y(k [ k 1) +
N
i=1
i
i
(k) (245)
Y(k [ k) = Y(k [ k 1) +
N
i=1
I
i
(k). (246)
These equations should immediately be compared to the considerably more complex mul
tiple sensor form of the Kalman lter update Equations.
4.2.4 The Hierarchical Information Filter
It is now shown how the information lter may be partitioned to provide a simple hierar
chical estimation architecture based rst on the communication of the information terms
i() and I() from sensor nodes to a common fusion center, and second on the communica
tion of partial informationstate estimates from nodes to a central assimilation point. The
latter case corresponds to the algorithm developed in [28]. In Section 2.2.6 it was demon
strated that, because of simple information summation, the loglikelihood form of Bayes
theorem can readily be mapped to a number of dierent architectural forms. For the same
reasons, the information additions in Equations 245 and 246 can also be distributed in a
simple manner.
H
1
H
n
H
2
{ i
1
(k), I
1
(k) }
z
1
(k)
z
2
(k)
z
N
(k)
Sensor
Sensor Models
{ i
n
(k), I
n
(k) }
{ i
2
(k), I
2
(k) }
k k1
Central Processor
y(kk1), Y(kk1)
^
y(kk), Y(kk)
^
Prediction
Figure 31: A hierarchical data fusion system where informationstate contributions are
calculated at each sensor node and transmitted to a central fusion center where a common
estimate is obtained by simple summation. All state predictions are undertaken at the
central processor.
One hierarchical architecture that employs this additive property is shown in Figure
31 (this is the information lter form of Figure 12). Each sensor incorporates a full
BAE SYSTEMS Proprietary Information, ASCTA use only 106
state model and takes observations according to Equation 235. They all calculate an
informationstate contribution from their observations in terms of i
i
(k) and I
i
(k). These
are then communicated to the fusion centre and are incorporated into the global estimate
through Equations 245 and 246. The informationstate prediction is generated centrally
using Equations 221 and 222 and the state estimate itself may be found at any stage from
x(i [ j) = Y
1
(i [ j) y(i [ j). To avoid communicating predictions to nodes, any validation
or data association should take place at the fusion center.
k k1
Central Processor
y(kk1), Y(kk1)
^
y(kk), Y(kk)
^
{ i
1
(k), I
1
(k) }
z
1
(k)
k k1
H
T
R
1
1 1 y
1
(kk), Y
1
(kk)
~ ~
y
1
(kk1), Y
1
(kk1) ^
Prediction
{ i
n
(k), I
n
(k) }
z
n
(k)
k k1
H
T
R
1
n n
Prediction
y
n
(kk1), Y
n
(kk1) ^
y
n
(kk), Y
n
(kk)
~
~
Sensor node n
Sensor node 1
Figure 32: A hierarchical data fusion system where tracks are formed locally and com
municated to a central fusion site. Each sensor node undertakes a prediction stage and
maintains a track based only on its local observations. The central processor fuses these
tracks.
A second hierarchical system which allows local tracks to be maintained at local sensor
sites, is shown in Figure 32 (this is the information lter form Figure 13). In this system,
each sensing node produces local informationstate estimates on the basis of its own obser
vations and communicates these back to a central fusion center where they are assimilated
to provide a global estimate of informationstate. Let y
i
( [ ) be the informationstate
estimate arrived at by each sensor site based only on its own observations and local
informationstate prediction y
i
(k [ k 1). This local estimate can be found from a local
form of Equations 226 and 227 as
y
i
(k [ k) = y
i
(k [ k 1) +i
i
(k) (247)
and
Y
i
(k [ k) = Y
i
(k [ k 1) +I
i
(k). (248)
These partial informationstate estimates are communicated to a fusion center where they
can be assimilated according to
y(k [ k) = y(k [ k 1) +
N
i=1
[ y
i
(k [ k) y
i
(k [ k 1)] (249)
BAE SYSTEMS Proprietary Information, ASCTA use only 107
and
Y(k [ k) = Y(k [ k 1) +
N
i=1
_
Y
i
(k [ k) Y
i
(k [ k 1)
_
. (250)
With the assumption that the local predictions at each node are the same as the prediction
produced by a central fusion center, it can be seen that these assimilation equations are
identical to Equations 245 and 246.
k k1
Central Processor
y(kk1), Y(kk1)
^
y(kk), Y(kk)
^
y(kk)[y
n
(kk)y
n
(kk1)]
~
~
y(kk)[y
1
(kk)y
1
(kk1)]
{ i
n
(k), I
n
(k) }
z
n
(k)
k k1
H
T
R
1
n n
Prediction
y
n
(kk1), Y
n
(kk1) ^
y
n
(kk), Y
n
(kk)
~
~
Sensor node n
{ i
1
(k), I
1
(k) }
z
1
(k)
k k1
H
T
R
1
1 1 y
1
(kk), Y
1
(kk)
~ ~
y
1
(kk1), Y
1
(kk1) ^
Prediction
Sensor node 1
Figure 33: A hierarchical data fusion system where tracks are formed locally and commu
nicated to a central fusion site, then track updates are communicated back to the sensor
nodes. Each sensor node undertakes a prediction stage and maintains a global track based
on the observations of all sensor nodes.
A third hierarchical system which allows global tracks to be maintained at local sensor
sites is shown in Figure 33. This architecture is similar to that shown in Figure 32 except
that once a global estimate is obtained at the central fusion site, it is communicated back
to the local site where it is assimilated to form a global estimate. The advantage of this
architecture is that each site can now act in an autonomous manner with access to
global track information.
There are a number of other possible implementations of these hierarchical estimation
equations, depending on where it is most ecient to generate predictions and where
dierent parts of the system model reside. The important point to note is that the
information lter provides a simple and natural method of mapping estimation equations
to dierent architectures.
4.2.5 The Decentralised Information Filter
It is a relatively simple step to decentralize the assimilation Equations 249 and 250 in
systems where there is a fully connected network of sensing nodes as shown in Figure
34. In this type of system, each node generates a prediction, takes an observation and
computes a local estimate which is communicated to all neighbouring nodes. Each node
receives all local estimates and implements a local form of the assimilation equations to
produce a global estimate of informationstate, equivalent to that obtained by a central
BAE SYSTEMS Proprietary Information, ASCTA use only 108
{ i
n
(k), I
n
(k) }
z
n
(k)
k k1
H
T
R
1
n n
Prediction
y
n
(kk1), Y
n
(kk1) ^
y
n
(kk), Y
n
(kk)
~
~
Sensor node n
~
[y
1
(kk)y
1
(kk1)]
{ i
1
(k), I
1
(k) }
z
1
(k)
H
T
R
1
1 1
Prediction
Sensor node 1
y
1
(kk1), Y
1
(kk1) ^
y
1
(kk), Y
1
(kk)
~ ~
k k1
{ i
2
(k), I
2
(k) }
z
2
(k)
k k1
H
T
R
1
2 2 y
2
(kk), Y
2
(kk)
~ ~
y
2
(kk1), Y
2
(kk1) ^
Prediction
Sensor node 2
~
[y
2
(kk)y
2
(kk1)]
~
[y
n
(kk)y
n
(kk1)]
Figure 34: A Decentralized, fully connected, data fusion system where tracks are formed
locally and communicated to neighbouring nodes where they are assimilated to provide
global track information. Each sensor node undertakes a prediction stage and maintains
a global track based on the observations of all sensor nodes. This architecture is also
directly equivalent to a broadcast or bus communication system.
fusion center. In eect, this is the same as replicating the central assimilation equations
of Figure 33 at every local sensor site, and then simplifying the resulting equations.
It is assumed that each local sensor site or node maintains a statespace model
(F(k), G(k), Q(k)) identical to an equivalent centralized model so that y
i
( [ ) y( [ )
for all i = 1, , N. Each node begins by computing a local estimate y
i
(k [ k) based
on a local prediction y
i
(k [ k 1) and the observed local information i
i
(k) according to
(Equations 221 and 222)
Prediction:
y
i
(k [ k 1) =
_
1
i
(k)G
T
(k)
_
F
T
(k) y
i
(k 1 [ k 1) +Y
i
(k [ k 1)B(k)u(k)
(251)
Y
i
(k [ k 1) = M
i
(k)
i
(k)
i
(k)
T
i
(k) (252)
where
M
i
(k) = F
T
(k)Y
i
(k 1 [ k 1)F
1
(k), (253)
i
(k) = M
i
(k)G(k)
1
i
(k), (254)
and
i
(k) =
_
G
T
(k)M
i
(k)G(k) +Q
1
(k)
_
. (255)
Estimate:
y
i
(k [ k) = y
i
(k [ k 1) +i
i
(k) (256)
Y
i
(k [ k) = Y
i
(k [ k 1) +I
i
(k). (257)
these partial informationstate estimates are then communicated to neighbouring nodes
where they are assimilated according to
BAE SYSTEMS Proprietary Information, ASCTA use only 109
Assimilate:
y
i
(k [ k) = y
i
(k [ k 1) +
N
j=1
[ y
j
(k [ k) y
j
(k [ k 1)] (258)
Y
i
(k [ k) = Y
i
(k [ k 1) +
N
j=1
_
Y
j
(k [ k) Y
j
(k [ k 1)
_
(259)
If each node begins with a common initial informationstate estimate y
j
(0 [ 0) = 0,
Y
j
(0 [ 0) = 0 and the network is fully connected, then the estimates obtained by each
node will be identical.
The quantities communicated between sensor nodes; ( y
j
(k [ k) y
j
(k [ k 1)) and
(
Y
j
(k [ k) Y
j
(k [ k 1)), consist of the dierence between the local information at
a time k and the prediction based only on information up to time k 1. This can be
interpreted as the new information obtained by that node at the current time step. Indeed,
the communicated terms are algebraically equivalent to i
j
(k) and I
j
(k); logically the new
information available at a time k is just the information obtained through observation
at that time. Thus, the operation of the sensor network can be envisioned as a group of
local estimators which communicate new, independent, information between each other
and which assimilate both local and communicated information to individual obtain a
globally optimal local estimate.
There are three interesting points that can be made about these decentralized equa
tions:
The additional computation required of each node to assimilate information from
adjacent nodes is small; a summation of vectors in Equation 258 and a summation of
matrices in Equation 259. This is a direct consequence of the use of the information
form of the Kalman lter which places the computational burden on the generation
of predictions.
The amount of communication that needs to take place is actually less than is
required in a hierarchical organization. This is because each node individually com
putes a global estimate so that there is no need for estimates or predictions to be
communicated prior to an estimation cycle. This results an a halving of required
communication bandwidth, which may be further improved if model distribution is
incorporated.
The assimilation equations are the same as those that would be obtained in a system
with distributed sensing nodes and a broadcast communication system.
The algorithm dened by Equations 251259 is appropriate for both fully connected sensor
networks or for sensors connected to a broadcast communication facility (such as a bus
or Blackboard).
BAE SYSTEMS Proprietary Information, ASCTA use only 110
4.3 Decentralised MultiTarget Tracking
4.3.1 Decentralised Data Association
Data association in distributed systems is a complex problem. The reason for this is
that hard association decisions made locally, in an optimal manner with respect to local
observations, may not be optimal at the global level when all sensor information is made
available. Further, an incorrect association decision is almost impossible to undo once
data has been fused into a track.
The normal approach to this problem is to maintain both a local and a global track
le and periodically to synchronize the two (see for example [11] pp602605). Alterna
tive approaches involve using either probabilistic data association methods or multiple
hypothesis trackers both of which avoid the need to make hard local association decisions.
All data association methods require that the normalised innovation is made available
at each of the local processor nodes. In a decentralised data fusion architecture the infor
mation transmitted from one node to another is in the form of the informationtheoretic
quantities i(k) and I(k). To implement a local validation procedure it is therefore neces
sary to derive an expression from the prediction and communication terms which allow
computation of a normalised innovation by every node on the basis of local information
y(k [ k 1) and Y(k [ k 1). The result is normalised information residual from which
an information gate, equivalent to the innovation gate, can be obtained.
To formulate the information gate, the inverse of the information matrix I(k) is re
quired. Generally, the dimension of the observation vector is less than that of the state
vector, so the information matrix I(k) = H
T
(k)R
1
(k)H(k) is singular since it has rank
n
z
, equal to the size of the observation vector, but has dimension n
x
n
x
, the size of the
state vector; and generally n
x
> n
z
. Consequently the inverse information matrix (the
corresponding covariance matrix) is not well dened and instead the generalisedinverse
I
(k) = E, (260)
where E is an idempotent matrix which acts as the identity matrix for the information
matrix and its generalisedinverse [31, 51]
I(k)E = I(k), I
(k)E = I
(k), (261)
so that
I(k)I
(k)I(k) = I(k), I
(k)I(k)I
(k) = I
(k). (262)
Amongst all possible generalised inverses that satisfy Equation 260, 261 and 262, the
most appropriate denition is that which projects I(k) into the observation space in the
following form;
I
(k)
= H
T
(k)
_
H(k)I(k)H
T
(k)
_
1
H(k) (263)
This generalised inverse exploits the role of H(k) as a projection operation, taking state
space to observation space, and H
T
(k) as a backprojection, taking observation space to
BAE SYSTEMS Proprietary Information, ASCTA use only 111
state space [56]. Both projections do not change the content of the projected matrix
so they can be applied without modifying the information contribution. Multiplying
Equation 263 by H(k)I(k) demonstrates this
H(k)I(k)I
(k) = H(k)I(k)H
T
(k)
_
H(k)I(k)H
T
(k)
_
1
H(k) = H(k) (264)
The innovation measure is essential for data association and fault detection. The
innovation is based on the dierence between a predicted and observed measurement
(k) = z(k) H(k) x(k [ k 1). (265)
In information form, the measurement information is provided by the information vector
i(k) = H
T
(k)R
1
(k)z(k). The information residual vector is therefore dened analogously
(k)
= H
T
(k)R
1
(k)(k), (266)
which is simply the innovation (k) projected into information space. Substituting Equa
tion 265 into Equation 266 gives
(k) = H
T
(k)R
1
(k)z(k) H
T
(k)R
1
(k)H(k) x(k [ k 1) (267)
or
(k) = i(k) I(k)Y
1
(k [ k 1) y(k [ k 1) (268)
The information residual variance is computed from
(k) = E(k)
T
(k) [ Z
k1
(k) . (269)
Substituting in Equation 266 gives
(k) = H
T
(k)R
1
(k)E(k)
T
(k) [ Z
k1
(k) R
1
(k)H(k)
= H
T
(k)R
1
(k)
_
H
T
(k)P(k [ k 1)H(k) +R(k)
_
R
1
(k)H(k)
= I(k) +I(k)Y
1
(k [ k 1)I(k)
= I(k)
_
I
(k) +Y
1
(k [ k 1)
_
1
I(k) (270)
The normalised information residual can now be computed from
(k) =
T
(k)
(k)(k) (271)
noting that the pseudoinverse for (k) is
(k) = H
T
(k)
_
H(k)(k)H
T
(k)
_
1
H(k)
= H
T
(k)
_
H(k)H
T
(k)R
1
(k)S(k)R
1
(k)H(k)H
T
(k)
_
1
H(k)
= H
T
(k)
_
H(k)H
T
(k)
_
1
R(k)S
1
(k)R(k)
_
H(k)H
T
(k)
_
1
H(k) (272)
BAE SYSTEMS Proprietary Information, ASCTA use only 112
and substituting Equations 266 and 272 into Equation 271 gives
T
(k)
(k)(k) =
T
(k)R
1
(k)H(k)H
T
(k)
_
H(k)H
T
(k)
_
1
R(k)
S
1
(k)R(k)
_
H(k)H
T
(k)
_
1
H(k)H
T
(k)R
1
(k)(k)
=
T
(k)S
1
(k)(k). (273)
That is, the normalised information residual is identically the conventional (observation)
residual.
Thus to implement a data association or gating policy for a decentralised sensing
network using the information lter, a normalised gate (or modied loglikelihood) can
be constructed using Equations 268 and 270. This gate will be identical to the gate used
in conventional multitarget tracking algorithms. Once the gate is established, it becomes
possible to use any of standard data association methods.
4.3.2 Decentralised Identication and Bayes Theorem
Decentralised data fusion principles can be easily extended to situations in which the
underlying probability densities are not Gaussian and indeed where the densities are
discrete. This provides a means of implementing decentralised (discrete) identication
algorithms. The method is based on the use of loglikelihoods and extends the hierarchical
loglikelihood architectures described in Section 2.2.6.
Recall the recursive form of Bayes theorem
P(x [ Z
k
) =
P(z(k) [ x)P(x [ Z
k1
)
P(z(k) [ Z
k1
)
, (274)
which may be written in terms of loglikelihoods as
ln P(x [ Z
k
) = ln P(x [ Z
k1
) + ln
P(z(k) [ x)
P(z(k) [ Z
k1
)
, (275)
where x is the state to be estimated and Z
k
is the set of observations up to the k
th
timestep. As has been previously demonstrated, the information form of the Kalman
lter, for example, can be derived directly from this expression.
In Equation 275, the term ln P(x [ Z
k1
) corresponds to information accumulated
about the state up to time k 1. The term
ln
P(z(k) [ x)
P(z(k) [ Z
k1
)
corresponds to the new information generated at time k. Equation 275 exactly represents
the communication requirements of a fully connected decentralised system, in which each
node communicates a likelihood based on its own observation to all others and receives the
BAE SYSTEMS Proprietary Information, ASCTA use only 113
z
i
(k)
k k1
log P( x{Z
k
}
i
)
Sensor i
log P( x{Z
k1
}
i
)
P(z
i
(k) x)
P(z
i
(k)Z
k
)
log
P(z
i
(k) x)
P(z
i
(k)Z
k
)
log
P(z
j
(k) x)
P(z
j
(k)Z
k
)
log
P(z
n
(k) x)
P(z
n
(k)Z
k
)
log
Communicated
Terms
Figure 35: A Decentralized, fully connected, Bayesian data fusion system appropriate for
decentralised identication tasks.
likelihoods of all its neighbours which are then fused to form a global estimate. Rewriting
Equation 275 in terms of an individual node i, this global estimate is computed as
ln P(x
i
[ Z
k
) = ln P(x
i
[ Z
k1
) +
j
ln
P(z
j
(k) [ x
j
)
P(z
j
(k) [ Z
k1
)
, (276)
where the summation represents the communicated terms. This is illustrated for a node
i in Figure 35.
BAE SYSTEMS Proprietary Information, ASCTA use only 114
5 Communication in Decentralised Sensing Systems
The issue of communication is a signicant issue in distributed and decentralised data
fusion systems. This is because of limited communication bandwidth, time delays and
communication failures that can occur between sensing and fusion processes. In decen
tralised systems, a signicant issue is also the potential for timevarying communication
topology.
This section considers the communication problem in depth. The fundamental issue of
determining what needs to be communicated is rst developed in terms of Bayes Theorem.
This is then translated in to both the information lter and discrete estimation form.
Using these basic algorithms, the issues of delayed and asequent data are then considered.
Algorithms are then developed for dealing with data that comes in out of time or out of
order. Finally, practical issues, including burst and intermittent communication, are
addressed in the design of communication channel algorithms.
5.1 Decentralised Communication Theory
The decentralised data fusion algorithms described so far are implicitly limited in requiring
full communication, either as a fully connected sensing network or as a broadcast system.
The reason for this is that it is assumed that all new information in the network is made
available to all sensors at observation time.
Fully connected, complete information networks are not practically viable: Any re
alisable distributed sensing network should be able to cater for a variety of communication
topologies, variable communication delays and insertion of new sensors into the network.
The key to providing these abilities lies in the algorithms used to decide what information
should be communicated between dierent sensing nodes. This is the focus of this section.
It is rst demonstrated that the need for fullyconnectedness is a result of the assump
tion that all nodes share a common prediction. Implementation of the straightforward
decentralised data fusion equations consequently results in a nearestneighbour commu
nication policy in which information is not propagated through the network.
To overcome this, the idea of a channel lter is introduced. A channel lter records
the information that is communicated between nodes. In a treeconnected network the
information communicated between nodes is clearly also the information they have in
common. By subtracting this common information from any future communication, only
new information is communicated. However, it is shown that, in general sensor network
topologies, the channel lter will not nd all common information. This is because the
common information between two nodes can not be established uniquely from what is
communicated down a channel. Three (suboptimal) methods of overcoming this problem
are described.
5.1.1 Communication Topologies
There are three basic communication topologies that are considered in this section;
BAE SYSTEMS Proprietary Information, ASCTA use only 115
Fullyconnected or broadcast topologies in which every node has direct access to
every other node in the network (Figure 36(a)). This type of topology occurs when
a bus or broadcast medium is used for communication. In fully connected arrange
ments, individual nodes may have intermittent connection to the communication
medium. However, it is generally assumed that once connected, they enjoy equal
access to other nodes information.
Treeconnected or singly connected topologies in which only one path exists be
tween any two nodes (Figures 36(b) and (c)). This type of topology rarely occurs
in physical practice. However, it can readily be imposed on a more general network
through the use of routing or tagging algorithms. In singlyconnected topologies,
nodes can again communicate intermittently. However, if a link is broken, then the
network will become disconnected if a dierent route can not be established. This
makes singlyconnected networks generally unreliable.
Arbitraryconnected or multiplyconnected topologies allow all nodes to connect
to any other node either through a single or, more generally multiple paths (Fig
ure 36(d)). In multiplyconnected topologies intermittent communication on any
one link will not necessarily halt the ow of information through other routes.
The dierent topologies determine dierent algorithms for communication. Fully con
nected and broadcast networks follow most simply from the algorithms previously de
veloped. However, channel lters (to be described) are often still required to deal with
intermittent, delayed and burst communications.
In more general networks, the key data fusion problem turns out to be how to to de
termine the common information between two nodes. This information must be removed
before two nodes can freely communicate with each other. In fully connected networks,
the common information is simply the information the two nodes had in common last time
they communicated. In tree or singlyconnected networks, there is also a simple means
of determining common information by simply remembering (practically, summing) the
information that has previously been communicated down the channel connecting two
nodes. This accommodates information that has been previously shared while allowing
new information, from further down the network, to be propagated.
In general networks however, there is no general optimal algorithm for both propa
gating information and maintaining optimality within the locality constraints of a de
centralised sensing network (the proof of this result can be found in [58]). However,
communication in arbitrary networks can be accommodated either by relaxing the con
straints imposed on decentralised systems, or by using suboptimal fusion algorithms. In
the former case, methods such as routing or tagging can be used to reduce the problem to
one of a singlyconnected network where channellter algorithms can be used to deter
mine common information. This approach is most appropriate for networks that already
have some stable structure; communication hierarchies, or sensors on a single platform.
However, in largescale networks in which the topology changes dynamically on a regular
BAE SYSTEMS Proprietary Information, ASCTA use only 116
basis, suboptimal methods of determining the common information are most appropriate.
This is typical of very large networks operating in tactical situations.
5.1.2 Bayesian Communication in Sensor Networks
Expressions for the common information between two nodes are now derived from Bayes
theorem. This in turn establishes the relationships between loglikelihoods from which
common information lters can be derived.
Consider a sensor node i. The information set, consisting of all locally available
observations, is denoted Z
i
. Let x dene the state to be estimated, then P(x [ Z
i
) denes
the posterior probability on the state given the locally available information. Now the
set Z
i
contains information from both observations made locally at node i as well as
information communicated to it from other nodes. When it interacts and exchanges
information with another node j, it is essential to understand what information results.
In general, if there are two nodes i and j, with information sets Z
i
and Z
j
respectively,
who engage in communication, it is desirable to construct the common information set
Z
i
Z
j
constructed from the union of the two local information sets. This may then be used
to construct a posterior distribution P(x [ Z
i
Z
j
) for the state based on both sources
of information. If the two information sets have no information in common Z
i
Z
j
= ,
then it is simple to construct the union of the sets from a sum. However if the two sets
do have information in common Z
i
Z
j
,= then the problem of constructing the union
Z
i
Z
j
resolves to nding the common information Z
i
Z
j
. Determining this common
information turns out to be key to the decentralised communication problem.
The most general solution to the common information problem is in the form of Bayes
Theorem. Consider the interaction of pairs of nodes. For each communicating pair (i, j),
the required probability is P(x [ Z
i
Z
j
). Let the union of the individual observation
information sets be partitioned into disjoint sets as
Z
i
Z
j
= Z
i\j
Z
j\i
Z
ij
(277)
where
Z
i\j
= Z
i
Z
ij
, Z
j\i
= Z
j
Z
ij
, Z
ij
= Z
i
Z
j
,
and where the notation pr (the restriction operation) means elements of the set p ex
cluding those elements that are also in set r. Note also that
Z
i\j
Z
ij
= Z
i
, Z
j\i
Z
ij
= Z
j
.
Then,
P(Z
i
Z
j
[ x) = P(Z
i\j
Z
j\i
Z
ij
[ x)
= P(Z
i\j
[ Z
j\i
Z
ij
, x)P(Z
j\i
Z
ij
[ x) = P(Z
i\j
[ Z
ij
, x)P(Z
j
[ x)
=
P(Z
i\j
Z
ij
[ x)
P(Z
ij
[ x)
P(Z
j
[ x) =
P(Z
i
[ x)
P(Z
ij
[ x)
P(Z
j
[ x)
=
P(Z
i
[ x)P(Z
j
[ x)
P(Z
i
Z
j
[ x)
. (278)
BAE SYSTEMS Proprietary Information, ASCTA use only 117
3 4 5 6 7 8
x 10
4
2
3
4
5
6
7
8
x 10
4
Platforms and Communication Links
xposition (m)
y
p
o
s
it
io
n
(
m
)
Sensors
Communication Links
(a)
2 3 4 5 6 7 8
x 10
4
2
3
4
5
6
7
8
x 10
4
Platforms and Communication Links
xposition (m)
y
p
o
s
it
io
n
(
m
)
Sensors
Communication Links
(b)
3 4 5 6 7 8
x 10
4
2
3
4
5
6
7
8
x 10
4
Platforms and Communication Links
xposition (m)
y
p
o
s
it
io
n
(
m
)
Sensors
Communication Links
(c)
2 3 4 5 6 7 8
x 10
4
1
2
3
4
5
6
7
8
x 10
4
Platforms and Communication Links
xposition (m)
y
p
o
s
it
io
n
(
m
)
Sensors
Communication Links
(d)
Figure 36: Dierent communication topologies for a 20 node network: (a) Fully connected;
(b) and (c) tree connected; (d) general network topology.
BAE SYSTEMS Proprietary Information, ASCTA use only 118
Substituting Equation 278 into Bayes theorem gives
P(x [ Z
i
Z
j
) =
P(Z
i
Z
j
[ x)P(x)
P(Z
i
Z
j
)
=
P(Z
i
[ x)P(Z
j
[ x)
P(Z
i
Z
j
[ x)
P(x)
P(Z
i
Z
j
)
=
P(x [ Z
i
)P(x [ Z
j
)
P(x [ Z
i
Z
j
)
P(Z
i
)P(Z
j
)
P(Z
i
Z
j
)
P(Z
i
Z
j
)
= c.
P(x [ Z
i
)P(x [ Z
j
)
P(x [ Z
i
Z
j
)
. (279)
This shows that the relation between the posterior probability in the unknown state given
information from both nodes, P(x [ Z
i
Z
j
), as a function of the posteriors based only
on locally available information, P(x [ Z
i
) and P(x [ Z
j
), and the information the two
nodes have in common P(x [ Z
i
Z
j
).
Taking logs of Equation 279, gives the intuitive result
ln P(x [ Z
i
Z
j
) = ln P(x [ Z
i
) + ln P(x [ Z
j
) ln P(x [ Z
i
Z
j
). (280)
Equation 280 simply states that the fused information is constructed from the sum of the
information from each of the nodes minus the information they have in common. The
term ln P(x [ Z
i
Z
j
) describes the common information between two nodes which must
be removed before fusion.
5.1.3 Identication of Redundant Information in Sensor Networks
The key step in deriving fusion equations for decentralised sensing networks is to identify
the common information Z
i
Z
j
between estimates so that it is not used redundantly.
The problem of accounting for redundant information in decentralised communication
structures is encountered in many applications
10
. In decentralised data fusion systems, the
incorporation of redundant information may lead to bias, overcondence and divergence
in estimates.
The problem of identifying common information is most generally considered in terms
of information sets. A neighbourhood of a node is the set of nodes to which it is linked
directly. A complete neighbourhood includes the node itself. A node i forms an infor
mation set Z
k
i
at time k based on a local sensor observation z
i
(k) and the information
communicated by its neighbours. The objective of each node is to form the union of the
information sets in the complete neighbourhood [N
i
] :
j[N
i
]
Z
k
j
. In particular, consider
communications between node i and a neighbour j. The union of information sets, on
which estimates are to be based, may be written as
Z
k
i
Z
k
j
= Z
k
i
+Z
k
j
Z
k
ij
, (281)
10
In human communication decentralised and cyclic communication structures give rise to rumour
propagation or the chicken licken problem.
BAE SYSTEMS Proprietary Information, ASCTA use only 119
that is, the union is equal to the sum of information sets communicated minus the inter
section of, or common information between, these information sets. A fully decentralised
solution to the estimation problem is only possible where the intersection or common
communicated information Z
k
ij
can be determined from information which is available
locally.
The topology of the sensing network is the most important factor in determining
common communicated information. Consider the following three cases:
Full Connection: Consider the case in which each node is connected to every other
in a fully connected, or completely connected topology (Figure 38). In this case, the
sensor nodes may acquire observation information from the entire network through
direct communication and the neighbourhood of any node is the full network. In a
fully connected network, the problem of locally detecting and eliminating redundant
information is considerably simplied as every node has access to the same infor
mation. In this situation, the estimates formed by the nodes are identical and new
information is immediately communicated after the removal of the estimate from
the previous timestep as
_
j[N
i
]
Z
k
j
=
_
_
j[N
i
]
Z
k
j
_
j[N
i
]
Z
k1
j
_
_
. (282)
This gives rise to a communication strategy where each node subtracts the estimate
formed at the previous timestep prior to communicating its current observation in
formation. Thus, Equation 282 is an informationset equivalent of Equations 299
and 300 which explicitly subtract the common prediction from the local partial
estimates. Since the fully connected condition means that global and local informa
tion are in fact the same, this must be regarded as a special case of the common
information problem.
Tree Connection: In a tree connected topology, there is only one path between
each pair of nodes (see Figures 39 and 40). Therefore, a receiving node can be
certain that the only redundant information communicated by a neighbour is the
information that they have exchanged in the past. Thus, the observation information
history that is required for a tree communication system extends only to the previous
timestep. The only common information between node i and a neighbour j at time
k is the information which they exchanged in the last time interval k 1. This
results in a pairwise communication algorithm. For a node i on link (i, j),
Z
k
i
Z
k
j
= Z
k
i
+Z
k
j
Z
k
ij
,
where, crucially, the intersection can be found from information sets at the previous
timestep,
Z
k1
ij
= Z
k1
i
Z
k1
j
. (283)
BAE SYSTEMS Proprietary Information, ASCTA use only 120
This generalises to a communication strategy over the neighbourhood as
_
j[N
i
]
Z
k
j
= Z
k
i
+
_
_
jN
i
Z
k
j
_
jN
i
Z
k1
j
_
_
, (284)
where [N
i
] is the complete neighbourhood, and i , N
i
.
Arbitrary Networks: In an arbitrary network, the connectedness of nodes is un
known and may be partially connected, or non fully connected nontrees. Removal of
common information in arbitrary network topologies is complex because the pattern
of communication varies from node to node, yet each node is required to implement
the same algorithm. Consider again two communicating nodes
Z
k
i
Z
k
j
= Z
k
i
+Z
k
j
Z
k
ij
.
In the arbitrary network case, the intersection term Z
k
ij
can not be simply deter
mined from past communication on a single link. In the tree case, the common
information term depends only on the observation information of i and j. In a
nontree network, the information common to i and j may contain terms from other
nodes outside of the neighbourhood. This is because multiple propagation paths
are possible. Figure 37 illustrates the information that is communicated to nodes
in three cases. The communicated information terms at i and j are denoted T
i
and
T
j
respectively. The information T
j
is integrated into the observation information
sets Z
j
upon arrival at j. This information must then be acquired by i through
some operation of union or intersect of information sets. To main the constraints
imposed by full decentralisation, it must be possible to eliminate common commu
nicated information terms between any pair of communicating nodes, on the basis
of local information only. In a fully connected network, the solution is immediate
since T
i
= T
j
at every timestep and Equation 282 follows. A tree network is par
titioned about any pair of connected nodes (i, j) such that the information T
i
held
in the subtree from i and that held in the subtree from j are disjoint: T
i
T
j
= .
Therefore, i acquires the terms from the subtree T
j
only through j. In an arbitrary
network, it may be possible for i to acquire the information known to j along other
routes. The problem is that the communicated terms are not necessarily disjoint,
therefore, each node must be able to determine T
i
T
j
locally. As shown in Figure
37, information in the region of intersection arrives at both i and j. This multi
ple propagation must be accounted for. The problem of arbitrary networks can be
alternately be considered as estimation in networks which admit multiple cycles.
Determination of the communication requirements for nonfully connected decentralised
networks therefore hinges on the extraction of common information.
5.1.4 Communication Channel Filters
Equation 280 serves as the basis for developing information communication policies for
nonfully connected sensor networks.
BAE SYSTEMS Proprietary Information, ASCTA use only 121
T
i
i j
T
j
Subtrees Connected
Through (i,j)
T
i
=T
j
i j
Full Connection
T
i
i j
T
j
Subtnetworks Connected
Through (i,j)
Figure 37: Communicated information sets in the three classes of topology.
BAE SYSTEMS Proprietary Information, ASCTA use only 122
The probability density functions in Equation 280 can represent four dierent infor
mation or Kalman lter estimates:
The local estimate at node i:
x
i
(k [ k)
= Ex(k) [ Z
k
i
with covariance P
i
(k [ k).
The local estimate at node j:
x
j
(k [ k)
= Ex(k) [ Z
k
j
with covariance P
j
(k [ k).
The estimate based on the union of all information possessed by nodes i and j (in
eect the global estimate):
x
ij
(k [ k)
= Ex(k) [ Z
k
i
Z
k
j
with covariance P
ij
(k [ k).
The estimate based on the common information between nodes i and j:
x
ij
(k [ k)
= Ex(k) [ Z
k
i
Z
k
j
with covariance P
ij
(k [ k).
Following Equation 230; Substituting Gaussian distributions for the probability density
functions in Equation 280 and taking natural logarithms immediately gives the informa
tion lter equivalent of Equation 280 as
y
ij
(k [ k) = y
i
(k [ k) + y
j
(k [ k) y
ij
(k [ k) (285)
Y
ij
(k [ k) =
Y
i
(k [ k) +
Y
j
(k [ k) Y
ij
(k [ k) (286)
where
y
i
(k [ k) = y
i
(k [ k 1) +i
i
(k), y
j
(k [ k) = y
j
(k [ k 1) +i
j
(k) (287)
Y
i
(k [ k) = Y
i
(k [ k 1) +I
i
(k),
Y
j
(k [ k) = Y
j
(k [ k 1) +I
j
(k) (288)
It remains to evaluate the common information terms
y
ij
(k [ k)
= y
ij
(k [ k), Y
ij
(k [ k)
= Y
ij
(k [ k).
There are three cases
BAE SYSTEMS Proprietary Information, ASCTA use only 123
Fully Connected: When the network is fully connected, the common information
between two nodes at a time k is exactly the information communicated up to time
k 1. This is simply the common prediction
y
ij
(k [ k) = y
i
(k [ k 1) = y
j
(k [ k 1)
Y
ij
(k [ k) = Y
i
(k [ k 1) = Y
j
(k [ k 1). (289)
Substitution of Equations 289, 287 and 288 into Equations 285 and 286, yields the
previously derived decentralised data fusion Equations 299 and 300.
Tree Connected: When there is only one pathway joining any two sensor nodes,
the common information between these nodes can be obtained locally by simply
adding up the information that has previously been communicated on the channel
connecting the two nodes. Equations 258 and 259 provide a recursive expression to
for the total information communicated between the two nodes as
y
ij
(k [ k) = y
ij
(k [ k 1)
+[ y
i
(k [ k) y
ij
(k [ k 1)]
+[ y
j
(k [ k) y
ij
(k [ k 1)]
= y
i
(k [ k) + y
j
(k [ k) y
ij
(k [ k 1) (290)
and
Y
ij
(k [ k) = Y
ij
(k [ k 1)
+[
Y
i
(k [ k) Y
ij
(k [ k 1)]
+[
Y
j
(k [ k) Y
ij
(k [ k 1)]
=
Y
i
(k [ k) +
Y
j
(k [ k) Y
ij
(k [ k 1) (291)
This estimate of common information replaces the prior information terms in Equa
tions 258 and 259. The local updates at each node remain unchanged
y
i
(k [ k) = y
i
(k [ k 1) +i
i
(k) (292)
Y
i
(k [ k) = Y
i
(k [ k 1) +I
i
(k) (293)
and the assimilation stage becomes
y
i
(k [ k) = y
i
(k [ k) +
jN
i
[ y
j
(k [ k) y
ji
(k [ k 1)] (294)
Y
i
(k [ k) =
Y
i
(k [ k) +
jN
i
_
Y
j
(k [ k) Y
ji
(k [ k 1)
_
. (295)
This lter is clearly symmetric, y
ij
( [ ) = y
ji
( [ ), as two nodes have the same
common information, so need only be computed once for each channel.
BAE SYSTEMS Proprietary Information, ASCTA use only 124
General Networks: In networks which admit multiple cycles, it is possible for
information from a node j to be communicated to a node i both directly through
the link (i, j) and indirectly through other nodes connected to both i and j. In such
cases, simply adding up information transmitted through the direct connection (i, j)
is not the same as determining the common information y
ij
(k [ k) and Y
ij
(k [ k)
between these two nodes. Indeed, in general it is not possible to determine this
common information on the basis of local information only. A number of alternative
approaches to fusion in general networks are considered in Section 5.1.6.
Equations 290 and 291 dene an information lter which estimates the common infor
mation between nodes. The lter is completed by addition of a corresponding prediction
stage as
y
ij
(k [ k 1) =
_
1
ij
(k)G
T
(k)
_
F
T
(k) y
ij
(k 1 [ k 1)
+Y
ij
(k [ k 1)B(k)u(k)
Y
ij
(k [ k 1) = M
ij
(k)
ij
(k)
ij
(k)
T
ij
(k) (296)
where
M
ij
(k) = F
T
(k)Y
ij
(k 1 [ k 1)F
1
(k),
ij
(k) = M
ij
(k)G(k)
1
ij
(k), (297)
and
ij
(k) =
_
G
T
(k)M
ij
(k)G(k) +Q
1
(k)
_
. (298)
The channel lter is simply an informationstate estimator which generates estimates on
the basis of information communicated through a channel joining two adjacent nodes.
The channel lter The eect of the channel lter is to provide information to nodes from
further aeld in the network with no extra communication cost. If the network is strictly
synchronous (all nodes cycle at the same speed) then the information arriving at a specic
node will be time delayed in proportion to the number of nodes through which it must
pass. This gives rise to a characteristic triangular wavefront in the information map
which describes the way information is used at any one node to obtain an estimate.
5.1.5 Fully Connected and Broadcast Sensor Networks
It is appropriate to make some comments about the fully connected or broadcast com
munication Equations.
Recall the assimilation equations
y
i
(k [ k) = y
i
(k [ k 1) +
j
[ y
j
(k [ k) y
j
(k [ k 1)]
= y
i
(k [ k 1) +
jN
i
i
j
(k) (299)
BAE SYSTEMS Proprietary Information, ASCTA use only 125
and
Y
i
(k [ k) = Y
i
(k [ k 1) +
j
_
Y
j
(k [ k) Y
j
(k [ k 1)
_
= Y
i
(k [ k 1) +
jN
i
I
j
(k). (300)
Equations 299 and 300 make it clear that the estimate arrived at locally by a node i
is based on the observation information i
j
(k) and I
j
(k) communicated to it by nodes j.
If node i only communicates with a local neighbourhood N
i
, a subset of the complete
network, then the estimate arrived at locally will be based only on this information and
will not be equivalent to a centralised estimate.
In eect, each time a new local estimate y
j
(k [ k) is obtained at a node j, the prior
information at this node y
j
(k [ k 1) is subtracted to provide the new information
to be communicated to a node i. The assumption here is that the prior information
y
j
(k [ k 1) at node j is the information that nodes i and j have in common up to time
k1. Subtracting this from the local estimate should then give the new information to be
communicated from node j to node i. In the fullyconnected case, the prior information
at node j is indeed the common information between nodes i and j and so only the
information i
j
(k) is communicated. In the nonfully connected case, the prior information
at node j includes not only information common to node i but also information from other
branches in the network. Subtracting this from the local estimate however, again results
in only the information i
j
(k) being communicated to node i. This is because it assumes
that node i already has the information communicated to node j through other branches
of the network. The net result of this is that nodes only exchange information in their
immediate neighbourhoods and do not propagate information from distant branches.
Figure 38 shows a fully connected network in which local estimates will be equal to
global estimates using only Equations 299 and 300 for assimilation. Figure 39 shows a
linear connected sensor network and Figure 40 shows a tree connected sensor network.
Here, the estimates arrived at by each node, using Equations 299 and 300 for assimilation,
will be based only on observations made by sensors in the immediate neighbourhood.
5.1.6 General Network Structures
Tree networks are restricted to having only a single path between any two nodes. Clearly
such networks are not robust as failure of any one channel or node will divide the network
into two noncommunicating halves. Networks which are to be made reliable must ensure
that there are multiple, redundant, paths between sensor nodes.
However, the algorithms described in the previous section will not produce consistent
informationstate estimates for networks with multiple paths. The reason for this is that
the channel lter y
ij
( [ ), which is intended to determine information common to nodes
i and j, does so only by summing up information communicated through the channel
linking these two nodes. The assumption is that there is only one path between two
neighbouring nodes so the information they have in common must have passed through the
BAE SYSTEMS Proprietary Information, ASCTA use only 126
Z
B
Z
D
Z
A
Z
C
C
A B
D
Figure 38: A fully connected sensor network consisting of four nodes, A, B, C, and D.
All nodes in this network can generate estimates, equal to a centralised estimate, using
only Equations 299 and 300.
Z
B
Z
C Z
D
Z
A
A
B
C D
Figure 39: A linear sensor network consisting of four nodes, A, B, C, and D. Nodes in
this network, using only Equations 299 and 300, can only generate estimates based on
information available in their immediate neighbourhood.
BAE SYSTEMS Proprietary Information, ASCTA use only 127
Z
A
Z
B
Z
C Z
D
A B
C D
Figure 40: A tree connected sensor network consisting of four nodes, A, B, C, and D.
Nodes in this network, using only Equations 299 and 300, generate estimates based on
information available in their immediate neighbourhood.
communication link connecting them. If information is communicated to both nodes i and
j directly by some third node then this information will clearly be common information
held by both i and j, but will not appear in the channel lter y
ij
( [ ) because it has not
passed through the channel between them.
There are three possible solutions to this problem
Data Tagging: An obvious solution is to tag information as passes through the
network so that nodes can determine which comments are new and which have
already been communicated through other links. The problem with this method is
that it will not scale well with large data sets and with large sensor networks.
Spanning Trees: A viable solution is to use an algorithm which dynamically
selects a minimal spanning tree for the network. In eect, this allows redundant
links to be articially broken at runtime so allowing treebased communication
algorithms to function in a consistent manner. Links can be articially broken by
setting y
ij
( [ ) = y
j
( [ ), ensuring no information is communicated across the cut.
The distributed BelmanFord algorithm is one algorithm which allows such a tree
to be constructed in a fully distributed or decentralised manner (it is commonly
used for internet routing). Overall system performance depends on reducing data
delays by choosing a topology with short paths between nodes. The best topology
can then be implemented by articially breaking the necessary links and using the
channel lter algorithm. The network remains scalable although reconguration
may be necessary if nodes are to be added or removed. This may, at rst, seem to
BAE SYSTEMS Proprietary Information, ASCTA use only 128
be a disadvantage but, should the network sustain damage to links or nodes such
a strategy will enable the surviving nodes to recongure to form a new topology,
bringing formerly disused links into service to bypass damaged nodes or links. Such
a system will be more robust than a purely tree connected system which will be
divided into two noncommunicating parts by the loss of a node or link.
Dynamic Determination of CrossInformation: A promising approach in
highly dynamic large scale networks is to use a local algorithm which attempts
to determine how the information in two incoming channels is correlated. One such
method is the covariance intersect lter described in Section 3.1.6. This computes
the relative alignment between information matrices and produces a conservative
local update based on the worstcase correlation between incoming messages. The
advantage of this method is that it is fully general and will work in any network
topology. The disadvantage with this method is that the estimates arrived at, while
consistent, are often very conservative. The determination of the common informa
tion between two nodes in an arbitrary dynamic network is the subject of current
research [44].
5.2 Delays and Timing: Temporal Propagation of Information
This section denes the general state model and the notation and procedure for propagat
ing state both forward and backward in time. This is necessary to enable information to
be correctly timealigned between dierent sensors in communication and also to provide
algorithms to deal with delayed or asequent information. Together these algorithms then
allow a capability of dealing with intermittent and burst communications.
5.2.1 System Denitions
For the sake of completeness, some standard denitions are repeated here.
Consider the standard linear continuoustime state space model
x(t) = F(t)x(t) +B(t)u(t) +G(t)v(t), (301)
where x(t) is the state of interest, u(t) the control input, v(t) the driving noise, and
F(t), B(t), G(t) matrices modeling eects of state, control and noise inputs respectively.
Equation 301 admits a well known closedform solution
x(t) = (t, t
0
)x(t
0
) +
_
t
t
0
(t, )B()u()d +
_
t
t
0
(t, )G()v()d (302)
where (, ) is the state transition matrix satisfying the matrix dierential equation
(t, t
0
) = F(t)(t, t
0
), (t
0
, t
0
) = 1. (303)
The state transition matrix has three important properties:
BAE SYSTEMS Proprietary Information, ASCTA use only 129
1. It is uniquely dened for all t, t
0
in [0, ].
2. (The semigroup property) (t
3
, t
1
) = (t
3
, t
2
)(t
2
, t
1
).
3. (t, t
0
) is non singular and
1
(t, t
0
)=(t
0
, t).
Specically when F(t) = F is a constant matrix, the state transition matrix is dependent
only on the time interval and is given by
(t, t
0
) = (t t
0
) = exp F(t t
0
), (304)
5.2.2 Forward Propagation of State
Dene an asynchronous discrete time set t = t
0
, t
1
, , t
k
, , and a state transition
matrix taking the state from a time slot t
k
to t
k+1
F
(k+1,k)
= (t
k+1
, t
k
). (305)
It is assumed that the control and noise inputs over the time period t
k
to t
k+1
can be
approximated as constant and equal to u(k) and v(k) so that with the denitions
B
(k+1,k)
=
_
t
k+1
t
k
(t
k+1
, )B()d
G
(k+1,k)
=
_
t
k+1
t
k
(t
k+1
, )G()d, (306)
Equation 302 becomes
x(t
k+1
) = F
(k+1,k)
x(t
k
) +B
(k+1,k)
u(k) +G
(k+1,k)
v(k), (307)
where the noise sequence v(k) is assumed zero mean and white with variance Q
k
. Equa
tion 307 describes the evolution of the state forward from a time t
k
to t
k+1
. Equation 307
can be generalised to describe propagation forward by n time slots as
x(t
k+n
) = F
(k+n,k)
x(t
k
) +
n
i=1
B
(k+i,k)
u(k +i 1) +
n
i=1
G
(k+i,k)
v(k +i 1). (308)
5.2.3 Backward Propagation of State
Starting with Equation 307, decrementing the time index and rearranging gives
x(t
k1
) = F
1
(k,k1)
x(t
k
) F
1
(k,k1)
B
(k,k1)
u(k 1) F
1
(k,k1)
G
(k,k1)
v(k 1).
Noting that
F
1
(k,k1)
= F
(k1,k)
= (t
k1
, t
k
)
BAE SYSTEMS Proprietary Information, ASCTA use only 130
so consequently
F
1
(k,k1)
B
(k,k1)
= (t
k1
, t
k
)
_
t
k
t
k1
(t
k
, )B()d
=
_
t
k
t
k1
(t
k1
, t
k
)(t
k
, )B()d
=
_
t
k
t
k1
(t
k1
, )B()d
=
_
t
k1
t
k
(t
k1
, )B()d
= B
(k1,k)
,
and similarly
F
1
(k,k1)
G
(k,k1)
= G
(k1,k)
,
gives
x(t
k1
) = F
(k1,k)
x(t
k
) +B
(k1,k)
u(k 1) +G
(k1,k)
v(k 1), (309)
describing the evolution of the state backward in time from t
k
to t
k1
. Equation 309 can
be generalised to describe propagation backward by m time slots as
x(t
km
) = F
(km,k)
x(t
k
) +
m
i=1
B
(ki,k)
u(k i) +
m
i=1
G
(ki,k)
v(k i). (310)
5.2.4 Notation for Time Invariant Systems
If the system F(t), B(t), G(t) is time invariant, then the state, control and noise tran
sition matrices depend only on the integration time t
n
t
m
. It is a signicant aid to
understanding to dene a simplied notation for the case of timeinvariant systems. At a
time t
k
, dene
F
k
= F
k+
= F
(k+1,k)
, B
k
= B
k+
= B
(k+1,k)
, G
k
= G
k+
= G
(k+1,k)
as the onestep forward transition matrices,
F
k+n
= F
(k+n,k)
, B
k+n
= B
(k+n,k)
, G
k+n
= G
(k+n,k)
as the nstep forward transition matrices,
F
k
= F
(k1,k)
, B
k
= B
(k1,k)
, G
k
= G
(k1,k)
as the onestep backward transition matrices, and
F
km
= F
(km,k)
, B
km
= B
(km,k)
, G
km
= G
(km,k)
as the mstep backward transition matrices. Note, because of the asynchronous onestep
time dierence
F
k
= F
(k+1,k)
,= F
1
k
= F
1
(k1,k)
= F
(k,k1)
BAE SYSTEMS Proprietary Information, ASCTA use only 131
in general.
Equation 307 is now written as
x(t
k+1
) = F
k
x(t
k
) +B
k
u(k) +G
k
v(k),
and Equation 309 as
x(t
k1
) = F
k
x(t
k
) +B
k
u(k 1) +G
k
v(k 1).
The remaining sections dene notation for state estimates, covariances, information
states, information matrices and the propagation of these both forward and backward in
time.
5.2.5 Denition of Estimates and Information States
Observations of the state are assumed to be made (asynchronously) according to a stan
dard linear model in the form
z(k) = H(k)x(t
k
) +w(k)
where w(k) is taken to be a zero mean white noise sequence with covariance R
k
. Dene
the set of all observations up to t
k
as
Z
k
= z(1), z(2), , z(k) .
The state estimate x(i [ j) and its variance P(i [ j) at time t
i
given observations up to
time t
j
are dened as
x(i [ j)
= Ex(t
i
) [ Z
k
, P(i [ j)
= E(x(t
i
) x(i [ j))(x(t
i
) x(i [ j))
T
[ Z
j
.
The information vector y(i [ j) and information matrix Y(i [ j) at time t
i
given observa
tions up to time t
j
are dened as
y(i [ j)
= P
1
(i [ j) x(i [ j), Y(i [ j)
= P
1
(i [ j).
5.2.6 Forward Time Propagation
The propagation of information from a time t
k
to a time t
k+1
is considered. For the
Kalman lter, state estimates and covariances are propagated according to
x(k + 1 [ k) = F
k
x(k [ k) +B
k
u(k)
P(k + 1 [ k) = F
k
P(k [ k)F
T
k
+G
k
Q
k
G
T
k
For the information lter propagation, dene
(k + 1 [ k) = G
T
k
M
(k+1,k)
G
k
+Q
1
k
, (311)
BAE SYSTEMS Proprietary Information, ASCTA use only 132
where
M
(k+1,k)
= F
T
(k+1,k)
Y(k [ k)F
1
(k+1,k)
= F
T
(k,k+1)
Y(k [ k)F
(k,k+1)
= M
k
is a noiseless propagation of information. The onestep forward propagation of the
information matrix is then
Y(k + 1 [ k) = M
k
M
k
G
k
1
(k + 1 [ k)G
T
k
M
k
(312)
and the onestep forward propagation of the information state
y(k + 1 [ k) =
_
1 M
k
G
k
1
(k + 1 [ k)G
T
k
_ _
F
T
k
y(k [ k) +M
k
B
k
u(k)
_
=
_
1 M
k
G
k
1
(k + 1 [ k)G
T
k
_
F
T
k
y(k [ k) +Y(k + 1 [ k)B
k
u(k).
(313)
The nstep forward propagation are best obtained through direct evaluation of F
k
, G
k
and B
k
over a combined time interval.
5.2.7 Backward Propagation
For convenience, the propagation of information from a time t
k+1
to a time t
k
is considered.
The propagation of state estimates and covariances follows according to
11
x(k 1 [ k) = F
k
x(k [ k) +B
k
u(k 1)
P(k 1 [ k) = F
k
P(k [ k)F
T
k
G
k
Q
k1
G
T
k
For the information lter propagation, dene
(k 1 [ k) = Q
1
k1
G
T
k
M
(k1,k)
G
k
, (314)
where
M
(k1,k)
= F
T
(k1,k)
Y(k [ k)F
1
(k1,k)
= F
T
(k,k1)
Y(k [ k)F
(k,k1)
= M
k
is a noiseless backpropagation of information. The onestep backward propagation of
the information matrix is
Y(k 1 [ k) = M
k
+M
k
G
k
1
(k 1 [ k)G
T
k
M
k
(315)
11
Note, the indices (k [ k) and (k 1 [ k) do not imply an update at time k; exactly the same
propagation equations would apply for any indices of the form (k [ k m) and (k 1 [ k m).
BAE SYSTEMS Proprietary Information, ASCTA use only 133
and the onestep backward propagation of the information state
y(k 1 [ k) =
_
1 +M
k
G
k
1
(k 1 [ k)G
T
k
_ _
F
T
k
y(k [ k) +M
k
B
k
u(k 1)
_
=
_
1 +M
k
G
k
1
(k 1 [ k)G
T
k
_
F
T
k
y(k [ k) +Y(k 1 [ k)B
k
u(k 1).
(316)
The mstep backward propagation is also best obtained through direct evaluation of
F
k
, G
k
and B
k
over a combined time interval.
5.2.8 Delayed Data
The state can be optimally updated at time k using delayed data from time k 1 using
the following
Y
+
(k + 1 [ k) = Y
(k + 1 [ k) +I
+
(k)
y
+
(k + 1 [ k) = y
(k + 1 [ k) +i
+
(k) (317)
I
+
(k) = M
I
k
+M
Y
k
G
k
_
G
T
k
M
Y
k
G
k
+Q
1
k
_
1
G
T
k
M
Y
k
_
M
Y
k
G
k
+M
I
k
G
k
_ _
G
T
k
M
Y
k
G
k
+G
T
k
M
I
(k)G
k
+Q
1
k
_
1
_
M
Y
k
G
k
+M
I
(k)G
k
_
T
= M
I
k
+M
Y
k
Y
(k + 1 [ k)
_
M
Y
k
G
k
+M
I
k
G
k
_ _
G
T
k
M
Y
k
G
k
+G
T
k
M
I
k
G
k
+Q
1
k
_
1
_
M
Y
k
G
k
+M
I
k
G
k
_
T
i
+
(k) = Y
(k + 1 [ k)
_
F
k
(Y(k [ k) +I(k))
1
(i(k) I(k)Y
1
(k [ k) y(k [ k))
_
+I
+
(k)
_
F
k
(Y(k [ k) +I(k))
1
( y(k [ k) +i(k)) +B
k
u(k)
_
(318)
where
M
Y
k
= F
T
k
Y(k [ k)F
k
M
I
k
= F
T
k
I(k)F
k
(319)
are the predictions of the information matrix and the observation information respectively
from time k1 to time k, Y
n
(k) = G
T
n
(k)M
n
(k)G
n
(k) +Q
1
n
(k)
and where the subscript n denotes that the associated transition matrices are evaluated
from the integrals in Equation 302 over the interval nT. Likewise, from Equation 313
the informationstate vector can be propagated forward according to
y(k +n [ k) =
_
1 M
n
(k)G
n
(k)
1
n
(k)G
T
n
(k)
_ _
F
T
(k)
_
n
y(k [ k)
+Y(k +n [ k)B
n
(k)u
n
(k). (323)
An equivalent expression can be obtained for propagating information backwards in time
as
Y(k n [ k) = M
n
(k) +M
n
(k)G
n
(k)
1
n
(k)G
T
n
(k)M
n
(k) (324)
where
M
n
(k) =
_
F
T
(k)
_
n
Y(k [ k) [F(k)]
n
,
n
(k) = Q
1
n
(k) G
T
n
(k)M
n
(k)G
n
(k)
and where the subscript n denotes that the associated transition matrices are evaluated
from the integrals in Equation 302 over the interval nT. Similarly, the information
state vector can be propagated backward according to
y(k n [ k) =
_
1 M
n
(k)G
n
(k)
1
n
(k)G
T
n
(k)
_ _
F
T
(k)
_
n
y(k [ k)
+Y(k n [ k)B
n
(k)u
n
(k). (325)
The delayed data problem can now be solved using the following algorithm:
1. Backpropagate the estimate Y(k [ k) to the time, k n, at which the information
I(k n) was obtained, using Equations 324 and 325.
BAE SYSTEMS Proprietary Information, ASCTA use only 135
2. Add the delayed data I(k n) to the estimate Y(k n [ k) in the normal manner.
3. Propagate the new fused estimate back to time k using Equations 322 and 323.
It can be seen (Equation 317) that the net eect of this algorithm is to produce an estimate
in the form
Y(k [ k) = Y
Y
(k [ k) +Y
I
(k [ k) +Y
Y I
(k [ k)
where Y
Y
(k [ k) is the estimate obtained without the delayed information, Y
I
(k [ k) is
the estimate obtained using only the delayed information, and Y
Y I
(k [ k) is a cross
information term (uniquely dened Y
Y
(k [ k) and Y
I
(k [ k)) describing the crosscorrelation
between the information states caused by propagation through a common process model.
It should be noted that Y
Y I
(k [ k) is additive, so the obvious approximation
Y(k [ k) = Y
Y
(k [ k) +Y
I
(k [ k)
is conservative.
5.3.2 Channel Algorithms
The channel lter is a conventional information lter used to maintain an estimate of
common data passed through a particular channel. A channel lter on node i connected to
node j maintains the common information vector y
ij
(k [ k) and the common information
matrix Y
ij
(k [ k).
The prediction equations (for both forward and backward propagation) used in the
channel lter are the same as Equations 322325 described above. For the update stage
a channel lter receives information estimates from other nodes. When this happens, the
channel lter predicts the received information to a local time horizon and then determines
the new information at that time. The new information at the channel lter at node i
when data arrives through the channel connected to node j is the information gain from
node j
m
j
(k) = y
j
(k [ k n) y
ij
(k [ k m)
M
j
(k) = Y
j
(k [ k n) Y
ij
(k [ k m) (326)
Note now that the time indices Y
j
(k [ k n) and Y
ij
(k [ k m) are dierent as the dif
ferent nodes are asynchronous and the two information sets can be predicted over dierent
time horizons. This information gain m(k) and M(k) is analogous to the observation in
formation vector i(k) and I(k), and is merely added in the update stage. The update for
the channel lter between nodes i and j when new information has been received from
node j can then be written as
y
ij
(k [ k) = y
ij
(k [ k m) +m
j
(k)
Y
ij
(k [ k) = Y
ij
(k [ k m) +M
j
(k) (327)
BAE SYSTEMS Proprietary Information, ASCTA use only 136
This can be further simplied by substituting Equation 326
y
ij
(k [ k) = y
ij
(k [ k m) +m
j
(k)
y
ij
(k [ k) = y
ij
(k [ k m) + y
j
(k [ k n) y
ij
(k [ k m)
y
ij
(k [ k) = y
j
(k [ k n) (328)
This reveals a trivial update stage where the channel lter merely overwrites the previous
state with the newer one. When a new information set arrives at node i from node j, the
channel lter is updated by
y
ij
(k [ k) = y
j
(k [ k n)
Y
ij
(k [ k) = Y
j
(k [ k n) (329)
Alternatively, if the new information at the channel lter of node i is from the local lter
at node i, the update becomes
y
ij
(k [ k) = y
i
(k [ k)
Y
ij
(k [ k) = Y
i
(k [ k) (330)
While the normal information lter update of Equation 327 is perfectly valid, implemen
tation in the form of Equation 329 and 330 is much simpler as there is no computation,
just the overwriting of the previous state.
Together, the channel update equations allow data delayed from neighbouring nodes to
be fused. This together with local assimilation equations permits burst communication of
estimates accumulated over a time period, in a manner that ensures no double counting of
information. This also means that the channel lter algorithms are robust to intermittent
communication failure.
5.4 The Communication Channel Algorithm
This section provides a practical algorithm for the general implementation of a commu
nication channel.
5.4.1 Management of Communication
All communication algorithms should be node rather than network centric. This is be
cause, locally, the structure of the network can change dynamically whereas the operation
of a node can be considered xed and known. Thus a node is considered as an isolated
entity which engages in communication intermittently and opportunistically rather than
on a xed basis. This view also rmly xes a modular nodecentric model of system struc
ture. This is shown diagramatically in Figure 41. Broadly, the node generates information
measures y
i
(k [ k) given observations made locally and information communicated to the
node up to time k. The node implements a local prediction stage to produce information
measure predictions y
i
(k [ k 1) given all local and communicated data up to time k 1.
BAE SYSTEMS Proprietary Information, ASCTA use only 137
i(k)
y
i
(kk)
~
Channel Filters
Prediction
Preprocess
Channel
Manager
Sensor Node
y
i
(kk)
^
y
i
(kk1)
^
y
i
(kk) y
iq
(kkn)
^ ~
y
i
(kk) y
ip
(kkn)
^ ~
y
pi
(kk), y
qi
(kk)
^ ^
P
Q
Figure 41: Algorithmic structure of a decentralised sensing node.
At this time, local observations produce local information measures i(k) on the basis of
local observations. The prediction and local information measures are combined, by sim
ple addition, into a total local information measure y
i
(k [ k) at time k. This measure
is handed down to the communication channels for subsequent communication to other
nodes in the decentralised network. Incoming information from other nodes y
ji
(k [ k) is
extracted from appropriate channels and is assimilated with the total local information by
simple addition. The result of this fusion is a locally available global information measure
y
i
(k [ k).
The communication channels exploit the associativity property of information mea
sures. The channels take the total local information y
i
(k [ k) and subtract out all informa
tion that has previously been communicated down the channel, y
ij
(k [ k), thus transmit
ting only new information obtained by node i since the last communication. Intuitively,
communicated data from node i thus consists only of information not previously transmit
ted to a node j; because common data has already been removed from the communication,
node j can simply assimilate incoming information measures by addition.
Channel lters have two important characteristics:
1. Incoming data from remote sensor nodes is assimilated by the local sensor node
before being communicated on to subsequent nodes. Therefore, no matter the num
ber of incoming messages, there is only a single outgoing message to each node.
Consequently, as the sensor network grows in size, the amount of information sent
down any one channel remains constant.
2. A channel lter compares what has been previously communicated with the total
local information at the node. Thus, if the operation of the channel is suspended,
the lter simply accumulates information in an additive fashion. When the channel
is reopened, the total accumulated information in the channel is communicated
in one single message. The consequences of this are manyfold; burst transmission
of accumulate data can be employed to substantially reduce communication band
width requirements (and indeed be used to manage communications); if a node is
disconnected from the communications network, it can be reintroduced and infor
mation synchronised in one step (the same applies to new nodes entering the system,
BAE SYSTEMS Proprietary Information, ASCTA use only 138
dynamic network changes and signal jamming).
The remainder of this section details a practical algorithm for the implementation of
a channel.
5.4.2 Structure of the Communication Algorithm
p q
N(p)
Y
p
(k1k1) Y
q
(k1k1)
(a)
p q
N(p)
Y
p
(kk1) Y
q
(kk1)
Y
pq
(kk1)
Y
p,N(p)
(kk1)
(b)
p q
N(p)
I
p
(k)
Y
p
(kk)
~
=Y
p
(kk1)+I
p
(k)
I
q
(k)
Y
q
(kk)
~
(c)
p
Y
q
(k)=Y
q
(kk)Y
pq
(kk1)
Y
p
(kk)+
j
Y
j
(k)
Y
p
(kk)=
~
Y
N(p)
(k)
(d)
Figure 42: Structure of the general communication channel algorithm: (a) Initial state
of the node p prior to establishing communication with node q; (b) establishing common
information between nodes p and q at a time k in a channel pq; (c) computation of local,
partial estimates at the two nodes; (d) calculation of information gains at channel inputs
and subsequent updating of local estimates.
Consider a node p and its neighbourhood N(p) of nodes shown in Figure 42. Node
p has a local information estimate Y
p
(k 1 [ k 1) at time k 1. The node has a
number of communication channels open with its neighbourhood j N(p). Node p now
decides to also communicate with node q through a new channel pq. It is not generally
known a priori if p and q have previously communicated or indeed have been in indirect
communication with each other through another branch of the network. The rst step in
the general communication algorithm is establish what information node p and q have in
common. This is done in three stages:
1. A separate pointtopoint channel, labeled pq, is established between the two nodes.
The current estimates at the two nodes Y
p
(k 1 [ k 1) and Y
q
(k 1 [ k 1) are
placed in the channel.
BAE SYSTEMS Proprietary Information, ASCTA use only 139
2. The two estimates are then timealigned using the information state prediction
or retrodiction Equations 312316 described in Section 5.2. The time alignment is
normally taken to some time k at which the next available observation information is
to occur. This yields process yields two predictions Y
p
(k [ k 1) and Y
q
(k [ k 1).
3. The common information Y
pq
(k [ k 1) between nodes p and q is then determined.
This information may be computed by any of the methods described in Section 5.4.4.
The result of this process is a channel between the two nodes with a measure of the com
mon information Y
pq
(k [ k 1) between them at a time k. Similar measures Y
pj
(k [ k 1)
are computed for every connected node j N(p) in the neighbourhood of p. This is shown
diagramatically in Figure 42(b).
The next step is to communicate new information or information gain between node p
and its neighbours. As a rst step, it is assumed that node p acquires some local obser
vation information I
p
(k) at time k and fuses this with its local prediction Y
p
(k [ k 1)
to produce a local, partial, estimate
Y
p
(k [ k) at time k as
Y
p
(k [ k) = Y
p
(k [ k 1) +I
p
(k). (331)
Other nodes in the neighbourhood may also produce a similar estimate using observation
data acquired since the last communication or otherwise based only on the prediction (so
Y
p
(k [ k) = Y
p
(k [ k 1)). This is shown diagramatically in Figure 42(c). It is important
here to reiterate the nodecentric form of the algorithm: The node timealigns common
information in a single pointtopoint communication channel. The channels themselves
may well act asynchronously with each other. The dierent nodes in the network will
certainly act asynchronously.
The nal step is to compute the information gain from the channel following commu
nication. The most robust method of doing this is to communicate the total informa
tion
Y
q
(k [ k) from a node q to node p and then to subtract the common information
Y
pq
(k [ k 1) in the input channel of node p. This allows for the possibility that com
munication may be halted or changed part way through a message: By performing the
information gain calculation at the input stage, there is no need to guarantee a commu
nication until the message actually arrives. The resulting computation performed in the
incoming channel is simply
Y
pq
(k [ k) =
Y
q
(k [ k) Y
pq
(k [ k 1). (332)
Finally, the local estimate is updated by simply adding information gains coming from
each channel as
Y
p
(k [ k) =
Y
p
(k [ k) +
jN(p)
Y
j
(k [ k). (333)
This calculation is may be performed asynchronously at each node and with respect to
each channel at a node. This update is shown diagramatically in Figure 42(d).
BAE SYSTEMS Proprietary Information, ASCTA use only 140
5.4.3 Information Flow
Qualitatively, the eect of the algorithm is to pass all information between two nodes,
subtracting out only the information considered to be common. Substituting Equation
331 into Equation 332 to obtain
Y
pq
(k [ k) = I
q
(k) +Y
q
(k [ k 1) Y
pq
(k [ k 1) (334)
shows that local measurement information is always communicated in its entirety. Further,
the term Y
q
(k [ k 1)Y
pq
(k [ k 1) captures the additional information at node q not
already obtained at (or in common with) node p. This may include observations previously
obtained by at node q but not yet communicated, or information communicated to q by
other nodes in the network neighbourhood N(q). In the latter case, the eect of this term
is to propagate information through the network, from node to node, at each stage being
incorporated into a single information gain.
5.4.4 Computing the Common Information
The simplest method of computing common information is to assume that the two nodes
have all prior information in common so Y
pq
(k [ k 1) = Y
p
(k [ k 1) = Y
q
(k [ k 1).
With this substituted in Equation 334, the information gain is simply the locally ob
served information Y
pq
(k [ k) = I
q
(k). This algorithm will be called the allcommon
communication method and would be used in a fully connected or broadcast structure.
If the channel pq is subject to delay or intermittent availability, the allcommon algo
rithm can be easily extended to communicate all information observed as a single burst
communication. In this case, the common prior Y
pq
(k [ k n) is set to the common
estimate Y
p
(k n [ k n) at the last communication time k n, appropriately predicted
forward to the new communication time k. The local estimates
Y
q
(k [ k) computed in
Equation 331 now include all observation information I
q
(k j), j = n, , 0 taken over
the interval k n to k (and again, appropriately predicted through to time k). The
computed information gain Y
pq
(k [ k) obtained from Equation 332 now consists of all
this observation information with the last common prior Y
p
(k n [ k n) properly sub
tracted. It should be clear that no matter what the time delay n or the number of
observations made, only a single information gain Y
pq
(k [ k) is computed and commu
nicated. Fundamentally, this is what a burst communication consists of. However, it is
also worth noting that the time propagation of observation information terms I
q
(k n)
over a time n generally results in a net loss of information through the addition of the
process uncertainty accumulated over the delay time.
BAE SYSTEMS Proprietary Information, ASCTA use only 141
6 Management in Decentralised Systems
Managing sensor systems involves making decisions and taking actions under uncertainty.
In decentralised data fusion systems, three types of decisions must be made:
1. Sensor Management: When and what to sense;
2. Communications Management: When and what to communicate;
3. Organisation: How should the network be structured and controlled.
These decisions are associated with the evaluation, acquisition and exchange of infor
mation. Consequently, they all derive from very similar principles. Further, they follow
quite naturally from the ideas of information already developed in decentralised data
fusion systems.
In this course, Bayesian decision theory is used as the underlying basis for these
decision making problems. Bayesian decision theory provides a comprehensive foundation
for making decisions under uncertainty ([10] is a denitive if challenging introduction
to Bayesian decision theory). Section 2 and subsequent discussions has also shown that
Bayesian methods also provide a powerful means of describing information. Thus Bayesian
decision theory is a most appropriate methodology for studying management problems in
decentralised sensor networks. Section 6.1 provides a brief and necessary introduction to
Bayesian Decision theory and particularly its application to data fusion problems. Section
6.3 describes the extension of singledecision maker ideas to multiple decision makers.
The extension of singleperson decision making to multipleperson decision making is
not trivial, particularly if the principles of fully decentralised decision making are to be
adhered to.
Together Sections 6.1 and 6.3 cover the underlying theory necessary for the study of
sensor management problems. Section 6.4 then describes the application of the decision
theoretic approach to the problem of sensor management. Sensor management focuses
on issues such as sensor cuing, sensortotarget assignment, hando and emission man
agement. The focus is on single sensor abilities to obtain information and to share this
information with other sensors without other constraint. The sensor management prob
lem admits some simple and direct solutions in decentralised form and has consequently
seen the most comprehensive development.
Section 6.5 considers the communication management problem. This focuses on prob
lems imposed by limited communication bandwidth and limited achievable communica
tion topologies in decentralised sensor networks. In particular, the problems are concerned
with who to communicate with, when to communicate, and what to communicate within
some established constraints. Although less work has been done in this area than for
the sensor management problem, the ideas of information embodied in decentralised data
fusion algorithms naturally lead to a wellformed group of information maximisation
communication algorithms.
Section 6.6 considers the problem of organisation in decentralised networks. This
focuses on specic issues in establishing communication topologies and also in broader
BAE SYSTEMS Proprietary Information, ASCTA use only 142
issues of sensortoplatform assignment and platform trajectory optimisation. This is the
least developed area of the decentralised management problem and is an area of current
and future research. It involves much broader ideas of optimal control in distributed
systems. However, it also admits useful information maximisation solutions which,
while not optimal, give useful and quantiable organisation algorithms.
6.1 Bayesian Decision Theory
This section briey summarises the main elements of normative Bayesian decision theory
as it applies to sensor and sensor network management.
6.1.1 Structure of the Problem
Decision making begins with an unknown state x A. The state may be either continuous
or discrete and may indeed constitute a trajectory in time. Knowledge about the state is
assumed to be encoded in a probability density function in the form of a prior P(x) or as
a posterior P(x [ Z
k
).
Given knowledge of the state x, some action a / is to be taken. The action may
be to determine the true value of x (an estimation problem), to take some more general
control action, or indeed to decide to take a further observation. The action may be
single, or it may consist of some time sequence of actions a = a
1
, a
2
, , a
n
.
The action a is selected with respected to a utility function U(a, x) or equivalently
12
a loss function L(a, x). Utility or loss functions are realvalued functions which assign a
unique value (respectively a loss) incurred if the action a is taken when the true value of
the state is x.
A utility or loss function has the eect of placing a preference ordering on possible
actions. Simply an optimal action is one that is most preferred. The choice of optimal
action is encoded in a decision rule . A decision rule is a function taking knowledge of
the state and computing an action as a =
p
(x). Practically, a decision rule is a function
of the probabilistic knowledge on the state, P(x), rather than the actual state x (which
is not normally known with precision). However, it is traditional to blur this distinction
in notation and talk about a decision
p
(x).
Knowledge of the state P(x), the set of actions a, the utility (or loss) function U(a, x),
and specication of the decision rule constitute all the required elements of the decision
making and management problem.
12
Engineers, being pessimists by nature, tend to use the term Loss to compare dierent decisions.
Economists, being optimists by necessity, use the term Utility. With the obvious sign change, the terms
can be used interchangeably and results derived with respect to one can always be interpreted in terms
of the other.
BAE SYSTEMS Proprietary Information, ASCTA use only 143
6.1.2 Utility or Loss
A loss function L(, ) (or corresponding utility function U(, )) is dened as a mapping
from pairs of states and actions to the real line;
L : A / 1, U : A / 1.
The interpretation of a loss function L(x, a) is that L is the loss incurred in taking the
action a when the true state of nature is x. Similarly, the interpretation of a utility
function U(x, a) is that U is the gain obtained in taking the action a when the true state
of nature is x. In principle, it should be the case that for a xed state of nature, both
utility and loss will induce a preference ordering on /. To ensure this, loss and utility
functions must obey three rules called the utility or rationality axioms which guarantee
a preference pattern. We state these rules here in terms of utility, following convention.
An analogous set of rules applies to loss functions. For xed x A:
1. Given any a
1
, a
2
/, either U(x, a
1
) < U(x, a
2
), U(x, a
1
) = U(x, a
2
), or U(x, a
1
) >
U(x, a
2
). That is, given any two actions we can assign real numbers which indicate
our preferred alternative.
2. If U(x, a
1
) < U(x, a
2
) and U(x, a
2
) < U(x, a
3
) then U(x, a
1
) < U(x, a
3
). That is,
if we prefer action a
2
to a
1
and action a
3
to a
2
, then we must prefer action a
3
to
action a
1
; the preference ordering is transitive.
3. If U(x, a
1
) < U(x, a
2
), then U(x, a
1
) + (1 )U(x, a
3
) < U(x, a
2
) + (1
)U(x, a
3
), for any 0 < < 1. That is, if a
2
is preferred to a
1
then, in a choice
between two random situations which are identical except that both a
1
and a
2
occur
with probability , the situation involving a
2
will be preferred.
A proof that these axioms imply the existence of a utility function can be found in [10].
It is well known and much discussed by economists that people do not always act
rationally according to the rationality axioms and that indeed they normally have con
siderable problems constructing any kind of consistent utility function by which to judge
decisions. The denition of rationality does not really concern us when dealing with a
data fusion system that consists of deterministic algorithms; we can always enforce the
denition of rationality chosen. What is of importance is the construction of the loss
function itself. It is, in principle, important to ensure that the loss (or utility) function
employed truly represents the value we place on dierent decisions.
6.1.3 Expected Utility or Loss
Except in case of perfect knowledge, a utility function in the form of U(x, a) is not very
useful because the true state of nature x A will not be known with precision and
so the true utility gained in taking an action will not be known. Rather, there will be
a probability distribution P(x) summarizing all the (probabilistic) information we have
about the state at the time of decision making. With this information, one natural
BAE SYSTEMS Proprietary Information, ASCTA use only 144
method of dening utility is as an expected utility (or Bayes expected utility) which for
continuousvalued state variables is simply
(a)
= EU(x, a) =
_
xA
U(x, a)P(x). (336)
Clearly, Bayes expected utility simply weights the utility gained by the probability of
occurrence (an average utility). More pessimistic approaches could also be used. For
example, we could ask what possible value of x could would incur the smallest utility
gain. This approach results in a minmax or gametheoretic decision algorithm.
6.1.4 Bayes Actions
It is most likely that probabilistic information concerning x is obtained after taking a
number of observations, in the form of a posterior distribution P(x [ Z
n
). An expected
utility (or loss) following an action a may then be dened with respect to a specied
posterior distribution in the true state, as
P(x [ Z
n
)
(a)
= E
P(x [ Z
n
)
U(x, a)
=
_
x
U(x, a) P(x [ Z
n
) dx. (337)
The Bayes action a
= arg max
a
P(x [ Z
n
)
(a). (338)
Many wellknown decision theoretic principles can be stated in terms of a Bayes action.
For example, in estimation problems, the action set is made equal to the set of possible
states of nature (/ = A). In particular, the MMSE estimate dened by
x = arg min
aA
_
x
(x a)
T
(x a)P(x [ Z
n
)dx
= arg min
aA
_
x
L(x, a)P(x [ Z
n
)dx, (339)
is clearly a Bayes action with respect to a squared error loss function dened by L(x, x) =
(x x)
T
(x x).
6.2 The Utility of Information
For the data fusion problems under consideration in this course, the goal is almost always
the maximisation of information available to the system. As was demonstrated in Section
BAE SYSTEMS Proprietary Information, ASCTA use only 145
2.3, the Shannon and Fisher measures provide a natural concept of information in terms
of the compactness of probability density functions. Further, these information measures
are made available directly in the decentralised data fusion algorithms developed in this
course. It is therefore logical to establish a relationship between the utility of sensor
decision making and these information quantities. F
In the data fusion arena, the action set a = a
1
, a
2
, , a
n
corresponds to the dif
ferent sensing congurations or possible sensing strategies available to the system. In
general, the eect of each possible action a
i
is to induce a posterior probability distribu
tion P
i
(x)on the state x. A probabilistic outcome
i
of an action a
i
is dened simply as
the posterior pdf for each possible sensing action
i
= P
i
(X [ a
i
), i = 1, , n (340)
Utility theory can now be used to encode the preferential structure of a sensor faced
with several choices of sensing action in the form of an ordered set
1
,
2
, . . . ,
m
) , such that
1
_
2
_ _
m
, (341)
where
1
is the most preferred outcome. Decisions are now made through a maximization
of the expected utilities for all possible actions. Following Equation 338, the optimal
action a
is dened through
a
= arg max
a
()
= arg max
a
E
x, a) E
U(x, a) = E
log P(x [ a)
= H
(x [ a) (347)
or for each action
E
i
U(x, a
i
) = E
i
log P(x [ a
i
)
= H
i
(x [ a
i
), i = 1, , n (348)
where the dependence on the specic action a
i
is made explicit. The ecacy of the choice
of utility as the loglikelihood is now clear: The expected utility is simply the negative
entropy or Shannon Information.
The Bayes action dened in Equation 338 now has a natural interpretation as the
action which maximises information. This then allows all the informationtheoretic tools
developed in Section 2.3 to be brought to bear on the sensor and communication man
agement problems.
BAE SYSTEMS Proprietary Information, ASCTA use only 147
6.2.2 Information Filter Metrics
In the information lter, Information is quantied through the information matrix Y(i [ j)
or inverse covariance P
1
(i [ j). It follows from Equation 51, that the posterior informa
tion following the i
th
sensing action at time k is given by
I(k, i)
=
1
2
log [(2e)
n
[ Y
i
(k [ k) []
=
1
2
[(2e)
n
[ Y(k [ k 1) +I
i
(k) [] . (349)
The possible information matrices, I
i
(k), result directly from the geometry of the sensing
process as is described in Example 22.
A second related measure is the mutual information gain from a sensing action. For
the information lter (from the denition in Equation 53), this is given by
I(k, i) =
1
2
log
_
[ Y(k [ k 1) +I
i
(k) [
[ Y(k [ k 1) [
_
(350)
An example of the use of this information measure, in the context of sensor management,
will be provided in Example 28.
These two information measures can be generalised and used in all three areas of sensor
decision making. In sensor management, Equations 349 and 350 can be used to decide on
the assignment of sensors to targets and in sensor hando and cuing. In communications,
the same equations can be used to determine which information it is most informative to
communicate.
6.2.3 Discrete State Metrics
For discrete states, the Shannon information is used directly The posterior information
about a set of states x = x
1
, x
2
, i simply given by
I(k)
= Elog P(x [ Z
k
) =
i
P(x
i
[ Z
k
) log
_
P(x
i
[ Z
k
)
_
, (351)
The mutual information (from Equation 53) is
I(k) = E
_
log
_
P(z(k) [ x)
P(z(k))
__
=
i
P(z(k) [ x)
P(z(k))
log
_
P(z(k) [ x)
P(z(k))
_
=
z
(x
i
) log [
z
(x
i
)] . (352)
As with the information lter metrics, the discrete metrics in Equations 351 and 352 can
be used for the purpose of both sensor management and communication optimisation.
BAE SYSTEMS Proprietary Information, ASCTA use only 148
6.2.4 Composite Metrics
In multiobjective optimisation problems, simultaneous localisation and identication of
targets for example, composite information utility metrics can be employed. This is
normally done in a simple linear combination, for example;
I
composite
(x) = C
1
I
identity
(x) +C
2
I
location
(x),
where C
1
and C
2
are arbitrary weights.
6.3 Decision Making with Multiple Information Sources
Decisionmaking with multiple sources of information is fundamentally more complex
than the problem of single source decision making. The essential reason for this is that
it is dicult to provide a consistent measure of utility or loss for each decision maker.
In particular, two basic problems exist. First, how can the utilities of two dierent deci
sion makers be compared unless there is a common measure of value. This is known as
the problem of interpersonal utility comparison. Second, should decision makers aim to
maximize a utility which expresses only local preferences, or should each decision maker
evaluate its actions with respect to some common or group utility function. In the liter
ature on the subject, no single agreed solution exists for these two problems. However,
for specic decision making problems in which concepts of utility may be simplied and
made precise, it is possible to arrive at consistent and useful solutions to both the utility
comparison and the group utility problems. This is generally the case for the decen
tralised data fusion problems studied here in which utility is dened globally in terms of
information.
6.3.1 The Super Bayesian
The simplest solution to the multipleperson decision making problem is to assume that
everyone can send information to a single decision maker or Super Bayesian. The task
of this decision maker is, in the rst instance, to combine probabilistic information from
all sources and then to make decisions based on the global posterior P(x [ Z
k
) in the form
a
= arg max
a
(P(x [ Z
k
), a)
= arg max
a
EU(x, a) [ Z
k
, (353)
where U(x, a) is a single group utility function. The solution in this case is well dened in
terms of classical Bayesian analysis. However, this Super Bayesian approach is generally
not applicable or appropriate in decentralised data fusion systems where no single decision
maker is generally permitted.
6.3.2 Multiple Bayesians
Consider a system consisting of local Bayesians each able to obtain probabilistic informa
tion which they can shares with all other Bayesians. Locally, each Bayesian can formulate
BAE SYSTEMS Proprietary Information, ASCTA use only 149
a posterior distribution P(x
i
[ Z
k
i
) and from this select a Bayes action with respect to a
local utility U
i
(x, a). Acting locally, a Bayesian would choose the action which maximises
local expected utility:
a
i
= arg max
a
(P(x
i
[ Z
k
i
), a)
= arg max
a
EU
i
(x
i
, a) [ i
k
Z
. (354)
The essential problem here is that each decision maker may end up choosing a dierent
action. There is no guarantee that the resulting set of actions will be optimal in a global
sense or indeed that the actions do not interfere with each other.
A general solution to these problems presents a considerable challenge. The trivial
case exists where the optimal action a
i
at each Bayesian i happens to be the same for
all the Bayesians and thus becomes the group action. However, in general these local
decisions will not the same.
In the general case, there are two generally recognised solutions to this problem. The
rst is to optimize a weighted sum of local utilities in the form
a
= arg max
a
j
w
j
[EU
j
(x
j
, a) c(j)] (355)
where 0 w
j
1 and
j
w
j
= 1. The value c(j) is known as decision maker js security
level. Equation 355 is known as a linear decision pool. A second solution is to optimize
a product of local utilities in the form
a
= arg max
a
j
[EU
j
(x
j
, a) c(j)] , (356)
This is known as the Nash product [43]. In an information context, the linear decision
pool is most appropriate as the global information will be the sum of locally available
information.
In many situations, it may be that the various decision makers do not come to a single
unied opinion on what the global decision should be. Such a situation occurs when
the local utility of a global decision is smaller than the value that would be obtained by
agreeing to disagree. This leads to a range of further issues in cooperative game playing
and bargaining which are not described further here.
6.3.3 Practical Bargaining Solutions
To apply Equation 356 or Equation 355 to compute an optimal action requires information
about every sensors utilities for each possible group action. In a decentralised system,
this would require each sensor node to acquire all other sensor nodes expected utilities.
In general this would require each node to communicate all its expected utilities to all
other nodes. This is clearly not practical.
However an approximate, and iteratively optimal solution, to this problem can also
be obtained Suppose each sensor, by considering each a
i
/, computes the following
BAE SYSTEMS Proprietary Information, ASCTA use only 150
ordered set i.e.
_
i
(a
i1
),
i
(a
i2
), . . . ,
i
(a
ir
)
_
,
where
i
(a
i1
) _
i
(a
i2
) _ _
i
(a
ir
), (357)
such that a
i1
is sensor is most preferred action and a
ir
is is least preferred action. The
communication required can be reduced if each sensor i communicates only a subset of
the expected utility set corresponding to the more preferred actions. For example, such
a subset can be chosen by considering only those actions which result in expected utilities
greater than or equal to is current expected utility i.e.
i
(a
l
) :
i
(a
l
) c(i), (358)
where c(i) is sensor is current utility at time of decisionmaking. Another possibility is
to consider only expected utilities above a given threshold value. The eect of this is to
reduce the size of the expected utility set considered and hence the number of possible
actions to be considered.
An iterative bargaining algorithm can be stated as follows:
1. Order preferences. Each sensor i computes its ordered expected utility set and, by impli
cation, the corresponding ordered preferred action set i.e.
_
i
(a
i1
), . . . ,
i
(a
ir
)
_
,
_
a
i1
, . . . , a
ir
_
. (359)
2. Communicate rst preferences. Each sensor i communicates its most preferred action
a
i1
and the corresponding expected utility
i
(a
i1
) to other sensors. In the event that there
are two actions with an identical expected utility then both the actions are communicated.
3. Compare rst preferences. Each sensor i then compares its own rst preference a
i1
with
other received rst preferences a
j1
, j ^, if all refer to the same action then a = a
i1
. In
practice this will be unlikely except for very simple management problems. If the communicated
preferences do not refer to the same action, then each sensor i must communicate its expected
utility corresponding to each received preferred action a
j1
, j ^.
4. Maximize on rst preferences. Each sensor is then able to compute the group expected
utility
B
c
(a
j1
) =
1
(a
j1
) +
2
(a
j1
) + +
N
(a
j1
), (360)
corresponding to the most preferred action a
j1
for each sensor j: j ^. From this,
each sensor can nd the most optimal solution based on all the sensors rst preferences by
maximizing over the group expected utilities corresponding to each sensors rst preference
actions i.e.
arg max
a
j1
B
c
(a
j1
).
This completes the rst iteration of the process.
5. Repeat and maximize on subsequent preferences.. i.e. next communicate and
compute B
c
(a
j2
), j ^ and so on for subsequent preferences.
The number of iterations required in the bargaining depends on the degree of opti
mality required in the actions taken.
BAE SYSTEMS Proprietary Information, ASCTA use only 151
6.4 Decentralised Sensor Management
Sensor management is generally considered to encompass three main decisions:
SensorTarget Assignment. It is important to manage and coordinate the as
signment of sensors to targets in order to use sensing resources eectively. This also
ensures that all the targets that need to be observed are covered in a manner which
is consistent with system goals. To be most eective, the management of sensor
target assignments must take into account the dynamic nature of the problem, such
as in the case of moving targets or moving sensor platforms, by continually reviewing
current assignments.
Sensor Cueing and Hando. When using sensors with limited elds of view it
is important to ensure that targets which may pass out of view are not lost. Hence,
it may become necessary to cue sensors into whose eld of view a target may be
entering. Cueing may be done in a cooperative manner e.g. when sensors, capable
of obtaining dierent information, cooperate to resolve ambiguity concerning a par
ticular target. Hando refers to the transfer of the observation of a target by one
sensor to another. As expected, sensor cueing and hando should be consistent
with system goals.
Mode management. Often sensors are capable of operating in several modes,
therefore, it becomes necessary to make decisions concerning the most appropriate
mode for a given situation. This is similar to the situation where there are several
physically diverse sensors available. In such cases it becomes necessary to manage
such diversity i.e. make decisions regarding the appropriate sensor or sensor mode
for a particular observation.
6.4.1 Sensor to Target Assignment
The application of the methods described in Section 6.3 to the sensor to target assignment
problem is now described through a number of examples.
Example 26
Consider the sensing scenario depicted in Figure 44. Each of three (decentralised)
sensors are able to observe three possible targets T = t
1
, t
2
, t
3
. Thus, all possible actions
are constrained by a 11 sensortarget assignment. In this case, the action set is dened
by the various possible sensortarget assignments. These can, for example, be
/ =
_
_
a
1
= (t
1
j; t
2
i; t
3
k),
a
2
= (t
1
i; t
2
k; t
3
j),
a
3
= (t
1
k; t
2
i; t
3
j),
a
4
= (t
1
j; t
2
k; t
3
i),
.
.
.
_
_
. (361)
Some of these sensortarget assignments are illustrated in Figure 44.
BAE SYSTEMS Proprietary Information, ASCTA use only 152
i
j
k
t
2 t
3
t
1
(a)
i
j
k
t
2 t
3
t
1
(b)
i
j
k
t
2 t
3
t
1
(c)
i
j
k
t
2 t
3
t
1
(d)
Figure 44: Some alternate sensor to target assignments for the problem described in
Example 26.
BAE SYSTEMS Proprietary Information, ASCTA use only 153
Using the actionoutcome association suggested by Equation 361, values are assigned
to each of the utilities for each of the sensors i, j and k. For ease of illustration, scaled
values of typical expected utilities are used:
B
c
a
1
: t
1
20
i; t
2
16
j; t
3
18
k; 54
a
2
: t
1
20
i; t
2
14
k; t
3
15
j; 49
a
3
: t
1
10
k; t
2
16
j; t
3
33
i; 59
a
4
: t
1
10
k; t
2
22
i; t
3
15
j; 47
a
5
: t
1
10
j; t
2
22
i; t
3
18
k; 50
a
6
: t
1
10
j; t
2
14
k; t
3
33
i; 57
(362)
where B
c
is the group expected utilities obtained from Equation 355 with a security level
of zero (c(j) = 0).
In Equation 362, sensor node i has its highest local utility when it is assigned to target
t
3
. This will occur if either the group action a
3
or a
6
is taken. In contrast, sensor node j
has its highest local utility when it is assigned to target t
2
, which will occur in the group
actions a
1
or a
3
is taken. However, there is a degree of conict in local decision making as
sensor node k also has its highest local utility when it is assigned to target t
3
; thus simply
selecting locally maximising actions will not be optimal as target t
1
would go unobserved.
In total, the the ordered action preferences for each sensor node are
i : (a
3
, a
6
), (a
4
, a
5
), (a
1
, a
2
))
j : (a
3
, a
1
), (a
4
, a
2
), (a
5
, a
6
))
k : (a
5
, a
1
), (a
6
, a
2
), (a
3
, a
4
)) ,
where each tuple represents a pair for which the preferences, that is, the expected utilities
are the same. It can be seen that by communicating only the expected utilities correspond
ing to their rst preference actions, the solution obtained from maximizing on these, that
is a
3
, is the same as the one obtained by maximizing over all possible actions. This is
simply because the global optimum is indeed one of the local optimums; a fortuitous result
for this particular example.
The iterative algorithm is now applied to this example. The rst iteration yields the
following; considering is most preferred actions a
i1
B
c
(a
i1
) = B
c
(a
3
) =
i
(a
3
) +
j
(a
3
) +
k
(a
3
) = 59
B
c
(a
i1
) = B
c
(a
6
) =
i
(a
6
) +
j
(a
6
) +
k
(a
6
) = 57,
considering js most preferred actions a
j1
B
c
(a
j1
) = B
c
(a
3
) (already computed)
B
c
(a
j1
) = B
c
(a
1
) =
i
(a
1
) +
j
(a
1
) +
k
(a
1
) = 54,
BAE SYSTEMS Proprietary Information, ASCTA use only 154
considering ks most preferred actions a
k1
B
c
(a
k1
) = B
c
(a
1
) (already computed)
B
c
(a
k1
) = B
c
(a
5
) =
i
(a
5
) +
j
(a
5
) +
k
(a
5
) = 50.
A second bargaining iteration yields
B
c
(a
i2
) = B
c
(a
5
) (already computed)
B
c
(a
i2
) = B
c
(a
4
) =
i
(a
4
) +
j
(a
4
) +
k
(a
4
) = 47,
considering js 2nd most preferred actions a
j2
B
c
(a
j2
) = B
c
(a
4
) (already computed)
B
c
(a
j2
) = B
c
(a
2
) =
i
(a
2
) +
j
(a
2
) +
k
(a
2
) = 49,
considering ks 2nd most preferred actions a
k2
B
c
(a
k2
) = B
c
(a
2
) (already computed)
B
c
(a
k2
) = B
c
(a
6
) (already computed).
Thus, a maximization based only on the rst iteration in this example yields the solution
a
c
= a
3
as before. This highlights the fact that for relatively simple management problems,
there is rarely a need to have more than a single iteration of the bargaining process.
For management problems where there are a large number of sensors and a large action
set, more than one iteration may be required as the following example illustrates:
Example 27
Consider a system of 4 sensor nodes making observations of targets such that there
are 8 distinct sensing congurations possible, that is, / = a
1
, . . . , a
8
. The following are
the ordered sensor expected utilities for each of the actions:
sensor i sensor j sensor k sensor l
a
1
20 a
5
40 a
3
20 a
7
35
a
2
15 a
6
35 a
4
18 a
2
30
a
3
10 a
1
20 a
5
16 a
3
10
a
4
5 a
2
15 a
6
14 a
5
5
a
5
4 a
8
10 a
7
12 a
6
4
a
6
3 a
4
5 a
8
10 a
4
3
a
7
2 a
3
2 a
1
8 a
8
2
a
8
1 a
7
1 a
2
6 a
1
1.
(363)
The 1st iteration yields
B
c
(a
i1
) = B
c
(a
1
) = 49
B
c
(a
j1
) = B
c
(a
5
) = 65 a
B
c
(a
k1
) = B
c
(a
3
) = 42
B
c
(a
l1
) = B
c
(a
7
) = 50.
BAE SYSTEMS Proprietary Information, ASCTA use only 155
Thus according to the rst iteration the optimal action is a
5
. However, proceeding to the
2nd iteration yields
B
c
(a
i2
) = B
c
(a
2
) = 66 a
B
c
(a
j2
) = B
c
(a
6
) = 56
B
c
(a
k2
) = B
c
(a
4
) = 31
B
c
(a
l2
) = B
c
(a
2
) (already computed),
showing that the optimal action is updated to a
2
. A third iteration yields actions that have
already been considered and similarly for the 4th iteration. In the 5th iteration all the
actions have been considered except a
8
which yields
B
c
(a
j5
) = B
c
(a
8
) = 30. (364)
These examples demonstrate the process of decision making or bargaining between
decentralised decision makers. The remaining sections focus more on the computation of
information utilities.
6.4.2 Sensor Hando and Cueing
The control of a number of dierent sensors to track, and possibly identify is a key
sensor management problem. The following example demonstrates the problem of co
operative tracking in a decentralised framework using information measures to evaluate
target choices.
Example 28
Three tracking sensors at xed locations x
s,i
= [x
s,i
, y
s,i
]
T
i = 1, 2, 3 make range and
bearing observations of three targets located at x
j
(k) = [x
j
(k), y
j
(k)]
T
j = 1, 2, 3 at time k.
Each sensor can only track one target. The discrete control action is the target assignment.
The control objective is to nd the 1 to 1 sensor to target mapping that maximises global
knowledge of the target states. Target dynamics are represented by a probabilistic constant
velocity model. The observation vector z
i
(k) = [r(k), (k)]
T
, is a function of the state of
the target being observed.
z
i
(k) = h(x(k)) +v
i
(k)
where v
i
(k) is taken to be a zeromean uncorrelated Gaussian sequence with variance,
Ev
i
(k)v
T
i
(k) = R =
_
r
0
0
_
The observation model is
h(x(k)) =
_
r(k)
(k)
_
=
_
_
_
(x(k) x
s
)
2
+ (y(k) y
s
)
2
arctan
_
x(k)x
s
y(k)y
s
_
_
_
BAE SYSTEMS Proprietary Information, ASCTA use only 156
The Jacobian with respect to target state is
H(x(k)) =
_
_
x(k)x
s
(x(k)x
s
)
2
+(y(k)y
s
)
2
y(k)y
s
(x(k)x
s
)
2
+(y(k)y
s
)
2
y(k)y
s
(x(k)x
s
)
2
+(y(k)y
s
)
2
x
s
x(k)
(x(k)x
s
)
2
+(y(k)y
s
)
2
_
_
=
_
sin((k)) cos((k))
1
r
cos((k))
1
r
sin((k))
_
The expected observation information for this sensor model is given by
I(k) = H
T
(k)R
1
H(k) =
_
_
sin
2
((k))
r
+
cos
2
((k))
r
2
sin((k)) cos((k))
sin((k)) cos((k))
r
2
sin((k)) cos((k))
sin((k)) cos((k))
r
2
cos
2
((k))
r
+
sin
2
((k))
r
2
_
_
Note, the determinant [ I(k) [ =
1
r
2
_
y
i,1
(k [ k)
y
i,2
(k [ k)
y
i,3
(k [ k)
_
_ , Y
i
(k [ k) =
_
_
Y
i,1
(k [ k) 0 0
0 Y
i,2
(k [ k) 0
0 0 Y
i,3
(k [ k)
_
_
With entropic information measure
i
i
(k) =
1
2
log
_
(2e)
12
[ Y
i
(k [ k) [
_
=
1
2
3
j=1
log
_
(2e)
4
[ Y
i,j
(k [ k) [
_
With its estimate of the target state after each prediction step, sensor i constructs the
expected information gain from observing target j.
I
i,j
(k) = H
T
( x
i,j
(k [ k 1), x
s,i
)R
1
H( x
i,j
(k [ k 1), x
s,i
)
This is communicated between all sensor nodes. Each node can then form a matrix of the
mutual information gain for each sensor target assignment.
I
i,j
(k) =
1
2
log
_
[ Y
i,j
(k [ k 1) +I
i,j
(k) [
[ Y
i,j
(k [ k 1) [
_
The utility for control action a
l
is
U
a
l
(k) =
3
n=1
I
i
n
,j
n
(k), Where a
l,n
= i
n
j
n
, a
l
_
_
a
1
= (1 1, 2 2, 3 3)
a
2
= (1 1, 2 3, 3 2)
a
3
= (1 2, 2 1, 3 3)
a
4
= (1 2, 2 3, 3 1)
a
5
= (1 3, 2 1, 3 2)
a
6
= (1 3, 2 2, 3 1)
_
_
The optimal action is selected by
a(k) = arg max
a
l
A
U
a
l
(k)
BAE SYSTEMS Proprietary Information, ASCTA use only 157
The sensors now make observation z
i
(k) of their allocated target and communicate the
observation information i
i
(k) = H
T
( x
i,j
(k [ k 1), x
s,i
)R
1
z
i
(k) to all nodes. The nodes
then update their local information by
y
i,j
(k [ k) = y
i,j
(k [ k 1) +i
j
(k),
Y
i,j
(k [ k) = Y
i,j
(k [ k 1) +I
j
(k).
Results from a solution to this example problem are shown in Figure 45. Figure 45
(b) shows that the value of observing a target is range dependent. As the targets move
the optimal group control action switches. This decision and communication structure
has provided a coordinated solution to the global control objective. Local prior and com
municated external information allows each node to arrive at a solution it believes is best
for the group. This example indicates that without this external information the resulting
control action would dier. There are points where greedy allocation based on individ
ual sensor to target mutual information gain is not the best group decision. A powerful
result of the communicated information is that combined with the observation model, each
sensor can determine the utility associated with observing a target. The decision to switch
target allocation is made based on this utility comparison without the sensors observing
their future target. This action is referred to as sensor hando and cueing.
It can be seen that the sensor nodes do not require any knowledge of the other sensors
location or characteristics. Everything required to select the optimal action is contained in
the local y
i
(k [ k), Y
i
(k [ k) and the communicated I
i,j
(k). This is an important property
of the decentralized information lter.
Note that in this case the local sensor knowledge is the true global information since all
information is communicated between the sensor nodes. If all information is not available
at each node this formulation is still valid. However, the local utility value associated with
an action will dier between nodes.
Optimal conguration of vehicle and sensor trajectories for feature tracking and locali
sation is a logical extension of this example. The reason is the sensor model. While highly
simplied, the range bearing model used is representative of a real world sensor. The com
bined observation information from multiple simultaneous observations is a function of
the relative range and bearing between the sensors, vehicles and features. The dynamics
of the information measures suggest an optimal system conguration.
6.5 Communications Management
The communications resource in a decentralised system provides a mechanism for plat
forms and sensors to exchange information. If the communications resource is constrained,
in bandwidth, latency or connectivity, then it must generally be managed.
The information metrics developed in this section can be used to control the ow
of information between sensor nodes. The subject of communications management was
BAE SYSTEMS Proprietary Information, ASCTA use only 158
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.2
0.4
0.6
0.8
1
Sensor 1 Sensor 2
Sensor 3
x
y
sensor /target locations
Target 1 Target 2
Target 3
true state
estimate
(a)
10 20 30 40 50 60 70 80 90
0.75
0.8
0.85
s
e
n
s
o
r
2
10 20 30 40 50 60 70 80 90
0.75
0.8
0.85
s
e
n
s
o
r
3
10 20 30 40 50 60 70 80 90
0.75
0.8
0.85
0.9
s
e
n
s
o
r
1
Sensor / target utility
Target 1
Target 2
Target 3
(b)
0 10 20 30 40 50 60 70 80 90 100
1
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
3
sensor / target assignment
t
a
r
g
e
t
in
d
e
x
Sensor 1
Sensor 2
Sensor 3
(c)
10 20 30 40 50 60 70 80 90
39
39.5
40
40.5
41
G
lo
b
a
l
I
n
f
o
r
m
a
t
io
n
10 20 30 40 50 60 70 80 90
13.5
13.55
13.6
13.65
13.7
13.75
T
a
r
g
e
t
I
n
f
o
r
m
a
t
io
n
Target 1
Target 2
Target 3
(d)
Figure 45: Information theoretic approach to a discrete sensor management task. (a)
Problem geography, (b) Information based Utility for each individual sensor to target
assignment, (c) Entropic target information,
(d) optimal group sensor to target assignment
BAE SYSTEMS Proprietary Information, ASCTA use only 159
studied in detail by Rob Deaves in his thesis [18] and is also reported in the paper [19].
Here we briey discuss the key points:
The value of information to be communicated can be measured using Equations 349
352.
These information measurements can be implemented as part of the communication
channel structure. This is illustrated in Figure 46.
Information on dierent tracks and other objects can be evaluated. Communications
can then be ranked by information value. strategies include:
Only communicate the most informative track information when bandwidth
becomes available.
Save up information in a channel until the total becomes suciently informative
(burst communication).
Using a measure of mutual information with respect to previously commu
nicated information, decide which node to connect to to communicate track
information.
a number of other strategies are also possible.
Generally, each node operates opportunistically, computing the mutual information
gains possible on each connected channel to determine when and what to commu
nicate.
In a limited medium, shared by a number of sensors, some element of the bargaining
solutions described in Section 6.3.3 would need to be introduced. This is still a
subject of current research.
6.6 Organisation and Control of Network Resource
The problem of organisation and control in decentralised networks has not yet received a
great deal of attention. The problem of optimal organisation was considered in Peter Hos
thesis [29]. This considered the problem of determining the connectivity arrangements in
largescale networks. Typical objectives included maximising the mean information levels,
maximising the minimum information level, and minimizing information deviation. The
network optimisation was performed using a distributed BellmanFord algorithm with the
information metrics of Equations 349352 serving as distance measures.
However, the problem of network connectivity and topology is only part of the overall
problem. There are also issues coupled to the placement and indeed the trajectory of
platform mounted sensors. This is illustrated in Example 29.
Control in decentralised systems is complex because of the potential coupling between
decision makers actions (two bearingonly sensors engaged in tracking for example). The
general problem is an area of current research and is not further discussed in this course.
BAE SYSTEMS Proprietary Information, ASCTA use only 160
Log
Manager
Message
Communications
Management
M, m
^
^
y, Y, P,
~
~
~
~
(transmitted) M, m
on Channels
1, .. , N1
(received)
Communications Module
Clock
To File Current Time #3
Channel N1
Channel 1
M, m (received)
Figure 46: Communications management process is used to control the function of channel
lters.
Example 29
Continuous Area Coverage Example Two vehicles i = 1, 2 are exploring a terrain
area T(x, y) dened on the (x, y) plane. The (twodimensional) trajectories for the i
th
vehicle is dened by x
i
= [x
i
(k), y
i
(k)]
T
, k [1, N] and shown in Figure 47. Each vehicle
makes observations z
i
(k) of the terrain according to
z
i
(k) = T(x, y) +v
i
(k)
where v
i
(k) is taken to be a zeromean uncorrelated Gaussian sequence with a variance
proportional to the range between the vehicle and terrain feature,
Ev
i
(k)v
T
i
(k) = R
T,i
(k) = R
_
(x
z
(k) x
i
(k))
2
+ (y
z
(k) y
i
(k))
2
,
where x
z
, y
z
are the true terrain locations being observed.
The vehicle trajectories are assumed known and the kinematics of the terrain are sta
tionary. It is required to generate estimates for the terrain T(x, y) over the whole surface.
In this case, the state transition and observation matrices are simply the identity matrix;
F(k) = 1 and H(k) = 1, so the prediction and update stages of the information lter
reduce to;
Prediction:
Y
T
(k [ k 1) = Y
T
(k 1 [ k 1)
Update:
Y
T
(k [ k) = Y
T
(k [ k 1) +
2
i=1
R
1
T,i
(k)
where the subscript T makes it clear that these information quantities of dened at every
point x, y on the terrain.
BAE SYSTEMS Proprietary Information, ASCTA use only 161
The posterior entropic information on T(x, y) and contained in the estimate is given
by
i
T
(k) =
1
2
log [2eY
T
(k [ k)]
The mutual information gain expected for an observation z
i
(k) of a terrain element T(x, y)
is
I
T
(x
i
(k)) =
1
2
log
_
Y
T
(k [ k 1) +R
1
T,i
(k)
Y
T
(k [ k 1)
_
This information measure is a function over T(x, y). A utility measure for an observation
made at x
i
(k) is
U(x
i
(k), k) =
_ _
I
T
(x
i
(k))dxdy
This can now be employed to generate a control metric to determine the trajectory of each
vehicle to maximize the total information over the whole area. The trajectory utility;
U(x
i
) =
N
k=1
U(x
i
(k), k)
Figure 47 shows snapshots of the information measures over time. The vehicles start
on opposite sides of the region. They travel at constant velocity over the indicated trajecto
ries. The last leg of the rst vehicles path overlaps the rst leg of the vehicles path. Plots
(a, d, g, j) show the entropic information over the area. (b, e, h, k) indicate the mutual
information gain for the current observation. (c, f, i, l) show the potential information
gain utility measure for an observation was made at location (x, y). This example is set
up to illustrate the information theoretic metrics. It highlights that the value of making
observations at a location is dependent on time and that the value of vehicle trajectories
are coupled. These metrics can be combined with other constraints, objectives and costs
to form an optimal area coverage control problem and solution.
Vehicle trajectories in this example are deliberately set up to illustrate the information
measures. In the proposed decentralised architecture the information obtained by each node
is propagated throughout the system. Local control actions are selected with the benet of
communicated or estimated external information. It can be seen that the mutual infor
mation measure decreases exponentially as observations are made. Hence, the value of
making observations from a position is coupled to time. A consequence of this probabilis
tic approach is exact knowledge of state can never be obtained. A wide range of possible
control objectives, minimum control eort, minimise maximum uncertainty in minimum
time uniformity across region, are possible. In this example information is never lost.
BAE SYSTEMS Proprietary Information, ASCTA use only 162
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0
2
4
6
8
2
x
y
1
In
fo
rm
a
tio
n
(a)
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0.1
0.2
0.3
0.4
0.5
0.6
0.7
2
x
y
1
M
u
tu
a
l In
fo
rm
a
tio
n
(b)
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0.5
1
1.5
2
2.5
3
3.5
4
4.5
2
x
y
1
In
fo
rm
a
tio
n
M
e
tric
(c)
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0
2
4
6
8
x
2
1
y
In
fo
rm
a
tio
n
(d)
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0.1
0.2
0.3
0.4
0.5
0.6
0.7
x
2
1
y
M
u
tu
a
l In
fo
rm
a
tio
n
(e)
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0.5
1
1.5
2
2.5
3
3.5
4
4.5
x
2
1
y
In
fo
rm
a
tio
n
M
e
tric
(f)
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0
2
4
6
8
1
x
2
y
In
fo
rm
a
tio
n
(g)
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0.1
0.2
0.3
0.4
0.5
0.6
0.7
1
x
2
y
M
u
tu
a
l In
fo
rm
a
tio
n
(h)
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0.5
1
1.5
2
2.5
3
3.5
4
4.5
1
x
2
y
In
fo
rm
a
tio
n
M
e
tric
(i)
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0
2
4
6
8
2
x
1
y
In
fo
rm
a
tio
n
(j)
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0.1
0.2
0.3
0.4
0.5
0.6
0.7
2
x
1
y
M
u
tu
a
l In
fo
rm
a
tio
n
(k)
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0.5
1
1.5
2
2.5
3
3.5
4
4.5
2
x
1
y
In
fo
rm
a
tio
n
M
e
tric
(l)
Figure 47: Snapshots of information measures for an area coverage example over time.
Plots (a, d, g, j) show information (as negative entropy) over the area, (b, e, h, k) indicate
the mutual information gain for the current observation. (c, f, i, l) show the potential
information gain if an observation was made at location (x, y)
BAE SYSTEMS Proprietary Information, ASCTA use only 163
7 Applications and Practical Considerations
This section provides a brief description of a number of key decentralised data fusion
projects. The objective is to describe some of the practical issues involved in developing
and implementing decentralised data fusion systems.
We begin by reviewing three past projects; SKIDS, ISSS, and OxNav. These served as
the basis for the original development of the various decentralised data fusion algorithms.
The current ANSER project is then described in some detail. Finally, we include some
discussion on largescale sensor networks characteristic of ground and combined airground
surveillance systems.
7.1 Historical Projects
7.1.1 SKIDS
In the SKIDS project, a fully decentralised surveillance system was implemented using
four cameras and a Transputer based architecture. The network was a fullyconnected
pointtopoint topology. The system was capable of tracking multiple targets (humans
and robots). A decentralised dataassociation algorithm was developed, but this was
superseded in later research and so is not described here. A decentralised identication
algorithm was also developed but this too was later superseded. The algorithms developed
for the SKIDS project are fully described in Bobby Raos D.Phil. thesis [49], and in the
papers [50, 48, 47]. The SKIDS demonstrator, which continued to be rened and operated
for almost 10 years, laid the basis for all subsequent work on decentralised data fusion.
7.1.2 ISSS
An essential limitation with the original decentralized Kalman lter algorithm is that it
requires the sensor network to be fully connected so ultimately limiting the size of any
realisable decentralized sensing system. The need for fully connectedness is a consequence
of the assumption that common information between two neighbouring sites is simply the
prior information they share.
This observation led to an analysis of information ow in decentralised sensing net
works. By introduction of an additional lter associated with each communication link,
it was shown that tree connected network topologies can also be supported by the decen
tralized Kalman lter algorithm. This is described in detail in Stewart Grimes D.Phil.
thesis [24] and in [25]. Channel lters, as they became known, also address a number of
key issues in data asynchronicity, communications management and network reliability.
Simukai Utetes D.Phil. thesis [58] showed that, within the constraints imposed by the
denition of a decentralized sensing network, it is not in general possible to construct a
set of lters which can provide consistent estimates across an arbitrary network topology.
To overcome this, decentralized routing algorithms were developed to enable construction
of tree connected networks from networks of arbitrary topology [57, 60, 61, 59, 62]. Such
algorithms maintain both consistency across the estimators in the network and satisfy the
BAE SYSTEMS Proprietary Information, ASCTA use only 164
constraints of locality and modularity of a decentralized system. Many of these ideas were
developed on a largescale demonstrator as part of the ISSS project. This demonstrator
consisted of a model process control system consisting of over 200 sensors linked to over
thirty purpose designed decentralized processing sites. The ISSS demonstrator allowed
online network reconguration and software imposed communication bandwidth limita
tions. The demonstrator was designed to show scalability of decentralised data fusion
algorithms to large numbers of sensors.
Further work on communications management has also been undertaken by Rob
Deaves of Sowerby research center [18, 19]. This work builds on the idea of a chan
nel lter and other work on information modeling [36] to manage communication between
sensing platforms. The work is developed in a military context.
7.2 OxNav
A second limitation of the original decentralized Kalman lter algorithm is that it requires
a complete system model to be maintained at every sensor site. As the size and complexity
of a system the grows the need for a global model at each site becomes prohibitive. The
problem of distributing system models across a sensor network was initially considered by
Berg [6, 8, 9, 7]. Starting with a central state model and with the local observation models
associated with each sensor node, it was shown that only a locally observable submodel
of the central state model is required at each site to ensure consistent estimates across a
network. It was also shown that the transformations of model from central to local sites
can be combined to provide a single transformation from one site to another which can
be implemented as part of the nodetonode communication mechanism.
An important contribution of this work was in the geometric interpretation of infor
mation and the consequent explanation as to why information measures, and not state
estimates, are uncorrelated (orthogonal) and thus why information fusion is associative.
This result was also exploited in the development of the information gate; an information
form of the equivalent data validation or innovation gate commonly used data association
algorithms [21, 22, 23]. Data validation is essential in practical data fusion problems in
providing a means of associating dierent internal models to observations made and in
rejecting observations that are considered to be outliers or spurious. In decentralized
systems validation must often take place remotely from the original observation. Under
standing how an observation taken at one site with only a partial local model of the overall
system compares to observations taken at another site with a dierent partial model is
essential in being able to associate and validate information across a decentralized sens
ing network. This work is also continuing in the development of high integrity navigation
systems [32].
A second consequence of the ability to perform model distribution is that it also allows
a connection to be made between the decentralized data fusion techniques and the eld
of decentralized control. In decentralized control, components of an overall system model
are distributed amongst a number of sites that both take observations of the world and
exert control over the environment. The extension of the decentralized and distributed
BAE SYSTEMS Proprietary Information, ASCTA use only 165
Kalman lter to problems in decentralized control is described in [41, 42, 40, 39, 38]. It
is shown how the model distribution results obtained for the decentralized Kalman lter
lead to both the design of a controller and to the design of a control structure which is
characteristic of the interconnections involved in the physical system to be controlled.
Drive System
Nodes
Sonar Sensing
Nodes
Figure 48: The OxNav Vehicle; a fully modular fully decentralised navigation and control
system
These ideas on control were implemented in the OxNav project aimed at demonstrat
ing fully decentralised and modular mobile robot navigation and control. This was a
particularly challenging project combining almost all aspects of the theory in a single
demonstration system and requiring the physical realization of a mechanical and elec
trically modular system, and demonstrates more clearly than any other application the
potential advantages to be gained from a decentralized approach to systems design. A
modular vehicle consisting of a number of standardized modular cages was designed (Fig
ure 48). Each cage contained a specic part of the overall vehicle function; drive unit,
BAE SYSTEMS Proprietary Information, ASCTA use only 166
sensor, power distribution, communication systems [13, 14]. Each cage contained a proces
sor, power and communication facilities, and all local software to implement the required
decentralized functions of that unit. There is no central unit or processor where informa
tion is combined or where control is coordinated. A wide range of dierent vehicle systems
were constructed from a small number of standardized cages, without the need to change
either hardware or software. The decentralized control system for the vehicle demon
strated that the design of local decentralized control algorithms for an individual driven
wheel unit allows the control of vehicles with any number of and kinematic conguration
of driven and steered wheels. The decentralized navigation system is also described. The
system employs a number of modular tracking sonar units. Each unit employs a model of
vehicle motion to track environment features to provide independent estimates of vehicle
location. The estimates are exchanged between sonar units to provide global navigation
information. Vehicle guidance was achieved through exchange of information between
vehicle drive units and sonar navigation sensors. The OxNav project won a number of
major awards for innovation and industrial collaboration.
7.2.1 Organisation and Management
All early work on decentralized data fusion relied on the simple algebraic manipulation of
a centralized estimation algorithm to obtain an equivalent decentralized estimator. The
work by Manyika [33, 36, 34, 35] provided a completely dierent approach to the decentral
ized data fusion problem based on an informationtheoretic model of sensor observation
and data assimilation. Informationtheoretic models are much more fundamental to the
data fusion problem than conventional state estimation techniques. In particular, the
work demonstrated that all of the decentralized algorithms derived to this point can in
fact be directly derived from information theoretic concepts once the problem has been
dened through Bayes theorem in probabilistic form. It was also demonstrated that the
quantities computed at sensor nodes in both the discrete and continuous lters may be de
scribed in terms of information quantities alone. Furthermore, the informationtheoretic
development of the problem demonstrates clearly why the decentralized lter is struc
turally and computationally a more natural formulation of the data fusion problem than
is a conventional multisensor lter.
The information formulation of the decentralised data fusion problem was rst demon
strated on the OxNav project; in particular for navigation and for sensor management.
Once the nature of decentralized sensing has been described in the form of information
measures it became possible to quantify the value of a particular observation and the value
of a sensor site within a sensing organization. If the value of an observation or site can be
measured, then a number of additional questions may be asked of the sensing network.
In particular, it becomes possible to address the problem of sensor management and net
work organization on the basis of information maximizing strategies. Sensor management
addresses the problem of how individual sensor sites should act so as to maximize the
amount of information available to the network.
The solution to the sensor management problem relies on the local measurement of
BAE SYSTEMS Proprietary Information, ASCTA use only 167
mutual information gain resulting from a particular sensing action, and through the global
maximization of this quantity by each individual sensor site. Sensor organization addresses
the problem of what structure the sensing network should take under specic quantiable
system requirements. Given a probabilistic model for each sensor site and the observations
made by each sensor, what network structure will provide, for example, the most robust
system or the most informative system [29]. The solution to the sensor organization
problem relies on a composite measure of the information generated by each node in
the system, accounting for the possible connectivity between sensor sites and the local
observability of dierent states at dierent nodes in the network.
The sensor management and organization problem demonstrate more clearly than any
other aspect of the decentralized data fusion problem the value of using information
theoretic quantities to describe both individual sensors and the overall structure of the
sensing network.
7.2.2 OxNav System and Demonstrations
The OxNav system shown in Figure 48 consists of a number of modular system compo
nents. The main sensors are tracking sonars as shown in Figure 49. These are able to lock
on to and track dierent types of features in an environment and to produce estimates
of relative feature location and identity (corner,line, etc). The main control elements are
driven and steered wheels coupled with spacer blocks of similar geometry. Figure 50 shows
an arrangement of nine such units.
Figure 49: A photograph of the tracking sonar used on the OxNav vehicle.
Each module, sensor or drive unit, contains a standard processor and processor moth
erboard which serve as the computational element of the node. Nodes are connected
to each other via a network of pointtopoint high speed serial lines. The interconnects
BAE SYSTEMS Proprietary Information, ASCTA use only 168
Figure 50: A photograph of a large OxNav vehicle showing how wheel modules are com
bined.
between nodes (which can be seen atop the vehicles in Figures 48 and 50), can be made
arbitrarily. There is no central processor on the vehicle.
The arrangement of sonars on the four corners of a standard vehicle is shown in
Figure 51. As the vehicle moves through the environment, the lead sonars scan and
lock onto new features. These features are tracked with respect to the estimated vehicle
location to provide both location and identity elements. As the vehicle moves through
the environment, these targets are handedo to the rearmounted sensors. These sensors
lock on and track these features, providing vehicle location data. The management of
the sensors, the cuing and hando of targets between sensors, is handled in a fully
decentralised manner using a composite information measure of identity and location for
the targets. The algorithm employed is based on that described in Section 6.3.3. A typical
run through an environment, with targets located and tracked, is shown in Figure 52.
7.3 ANSER
7.3.1 Objectives
The primary objective of the ANSER project is to demonstrate fully decentralised and
modular picture compilation and terrain navigation on single and multiple ight platforms
(UAVs). In turn, this will be used to demonstrate, in a highly relevant form to BAE
Systems operating divisions, how decentralised systems architectures result in:
1. Modular packaging of sensor and data fusion algorithms.
2. Scalability and online exibility to addition of single and multiple sensors and ight
systems.
BAE SYSTEMS Proprietary Information, ASCTA use only 169
Figure 51: Arrangement of tracking sonar sensors on the OxNav vehicle. Typically, the
leadin sensors are used to acquire new navigation targets and the rear sensors are used
to track known targets for localisation. Management of sensors is handled through the
use of information metrics communicated between nodes.
3. Robustness and fault tolerance to failure in sensors, systems and platforms.
4. Providing a quantitative means of analysing systems performance issues such as
communication and integrity.
5. Increasing degrees of controllability and autonomy in sensors and ight systems.
These objectives are being met by developing, from the outset, fully modular navigation
and picture compilation instrumentation following the theoretical and practical principles
developed in previous collaborative research projects. This species logical packaging
of sensors with processors where possible, and explicit separation of functions such as
observation preprocessing, data assimilation and intermodule communication.
7.3.2 Demonstrations
Fundamentally, ANSER is a demonstration project: The objective is to demonstrate
functionality of decentralised data fusion theory and algorithms developed over the past
10 years in a form which is of direct relevance to BAE Systems business units.
The ANSER project calls for the simultaneous deployment of up to four UAVs in
decentralised conguration (see Figures 53 and 57). Four platforms are the minimum
allowing demonstration of nontrivial decentralised communication policies. Each UAV is
BAE SYSTEMS Proprietary Information, ASCTA use only 170
Figure 52: A run of the OxNav vehicle along a corridor showing dierent targets being
acquired and tracked.
BAE SYSTEMS Proprietary Information, ASCTA use only 171
Figure 53: The Mark I ANSER UAV.
Figure 54: The Mark III UAV showing nose housing for radar.
BAE SYSTEMS Proprietary Information, ASCTA use only 172
Figure 55: A further view of the Mark III UAV.
Figure 56: A further view of the Mark III UAV.
BAE SYSTEMS Proprietary Information, ASCTA use only 173
Vehicle Bus (CAN)
Flight
Sensors
GPS/IMU
GPS/IMU
and Flight
Controller
Flight Mode
Switch
Actuators
AirGround
Modem
Environment Bus (CAN)
Radar Node
AirtoAir
Communications
Radar
Laser Node Vision Node
Camera Laser Scanner
Figure 57: Functional architecture for the UAV. The internal structure provides a CAN
bus for vehicle functions (ight critical), and a CAN bus for payload and map information.
Each payload is fully modular and interchangeable.
equipped with inertial, GPS and ight data sensors. Each UAV is also equipped with two
payload sensors. The rst payload sensor is always a passive vision system. The second
payload is chosen from either a laser or a mmwave radar. Both the laser and the radar
are mechanically scanned using a common scanner design (see Functional Specication
documents for further details). The UAVs are own at the ACFR ight test facilities at
Marulin, 175Km south of Sydney.
The UAVs will use the payloads to perform two primary functions:
1. Picture compilation: The detection and tracking of multiple ground targets given
known locations (derived from GPS/IMU) for the ight platforms.
2. Simultaneous localisation and mapbuilding (SLAM): The detection and tracking of
multiple terrain features together with the simultaneous use of these in estimation of
platform location, without independent knowledge of platform location (no GPS).
Each function will be developed and demonstrated in fully decentralised and modular
form; across payloads on any one ight platform and across multiple payloads on multiple
ight platforms. Figure 58 shows the structure of onboard algorithms.
The scenarios to be own are aimed at demonstration of the key elements of the
decentralised data fusion method. The scenarios can be broken down in to four main
groups (see scenario denition documentation for details):
1. Single platform picture compilation: A single platform will be own with multi
ple payloads in picture compilation mode. Demonstrations include the modular
exchange of payloads (modularity, interoperability), and the inight failure and
reconguration of payloads (survivability and exibility).
BAE SYSTEMS Proprietary Information, ASCTA use only 174
Figure 58: Algorithmic architecture for the UAV. Each payload operates an independent
decentralised lter and corresponding channel lters for both picture compilation and
SLAM modes. The internal function of each sensor node is hidden and their operation is
transparent to the operation and location of other payloads.
2. Single platform SLAM: A single platform will be own with multiple payloads in
SLAM mode. Demonstrations include ight proving the SLAM method (not dis
cussed in this document), generation of terrain maps from decentralised payloads,
with the same modularity and payload reconguration abilities as are demonstrated
for picture compilation functions.
3. Multiple platform picture compilation: Multiple ight platforms will be own with
multiple payloads in picture compilation mode. Demonstrations will include the
extension of function from one to four platforms (scalability), the transparent use of
sensors on one platform by assimilation processes on other platforms (modularity,
interoperability), reconguration of sensing due to failure of individual payloads,
and task reconguration due to failure of complete ight platforms (survivability
and exibility).
4. Multiple platform SLAM: Multiple ight platforms will be own in SLAM mode.
Demonstrations will include the ability to share terrain maps between payloads on
dierent ight platforms, to fuse maps from geographically separated payloads, and
to demonstrate scalability and robustness of decentralised SLAM methods.
The scenarios form a logical progression for the implementation of the theory and demon
stration of algorithms.
BAE SYSTEMS Proprietary Information, ASCTA use only 175
7.3.3 Research
The ANSER project is focused on demonstration of existing theory. However, the com
plexity of the demonstrator still requires research in three main areas:
1. The extension of theory and methods to airborne scenarios: There is a necessary
process of mapping a general theory and set of methods to this specic, and demand
ing, application. Most previous work undertaken in decentralised methods has been
done on groundbased sensors or land vehicles. An air scenario has more degrees of
freedom, faster data rates, and larger demands on processing and communication
management.
2. The development of sensing, terrain data acquisition, and terrain or target repre
sentation methods: The development and packaging of payload sensors for airborne
application has been a signicant issue; particularly weight, volume and data acqui
sition speed. The extraction of appropriate terrain features for picture compilation
and SLAM, the modeling and communication of these between dierent payloads
has also been an area of signicant study.
3. The development of decentralised and informationtheoretic SLAM methods: One
signicant theoretical advance made in this project is the development of a decen
tralised formulation of the map building and SLAM problem. This has required
the reformulation of map building equations in informationtheoretic form and the
study of how maps from dierent platforms can be exchanged and assimilated.
ANSER has also provoked a number of new research areas, beyond the scope of the current
project, but of considerable future value in decentralised systems.
7.4 Implementation of DDF Algorithms in ANSER
This section describes the decentralised data fusion algorithms implemented in the ANSER
system. The approach is to present generic equations for the decentralised data fusion
problem and subsequently (in Sections 7.4.7 and 7.4.8 to instantiate process and observa
tion models specic to the dierent demonstrations and sensor nodes employed. Figure 59
shows the general structure of pointtopoint communications in the decentralised data
fusion architecture. This is generally what is implemented within the ANSER system.
On a single vehicle, this is practically mapped to a bus or broadcast network as shown
in Figure 60. However, this does not change the essential decentralised algorithms for
fusion and communication.
All the decentralised data fusion functions to be demonstrated in ANSER are contin
uous state estimation problems. The decentralised data fusion algorithms therefore all
employ the information lter for continuous state estimation. The more general discrete
state decentralised estimation problem employs the loglikelihood form of Bayes theorem
from which the information lter is derived. State estimates (tracks or landmark esti
mates) are therefore all described in terms of an information matrix Y
p
(i [ j) and an
information state y
p
(i [ j).
BAE SYSTEMS Proprietary Information, ASCTA use only 176
Sensor Node
Sensor
Fusion Processor
Communications Medium
Figure 59: The general structure of a decentralised data fusion system employing point
topoint communications.
With these denitions, the overall structure of a decentralised sensing node is shown
in Figure 61. The node executes four main functions:
1. Information state prediction, generating y(k [ k 1) and Y(k [ k 1) at the obser
vation time (generally asynchronous) from past information state values.
2. Observation preprocessing to provide the observation information vector and matrix
i(k) and I(k). If necessary these need to be associated with correct predictions, but
then fusion is a simple process of summation.
3. Information transmission through the channel lter. This requires the calculation
of new information to be communicated to connected nodes.
4. Assimilation of incoming data from channels. This generally requires temporal
alignment and compensation of delayed data eects.
State estimation takes place at the local lter, which simply sums the total information
from observation, communication and prediction. All communications to the network
pass through the channel manager, which handles timing and interfacing with the channel
lters. The channel lters keep an estimate of all information that has passed down a
particular channel and physically connects with the communications medium.
The main node operations are now described:
BAE SYSTEMS Proprietary Information, ASCTA use only 177
Communications Medium
Sensor Node
Sensor
Fusion Processor
Figure 60: General decentralised data fusion system mapped to a busbased communica
tions architecture within a single platform.
i(k)
y
i
(kk)
~
Channel Filters
Prediction
Preprocess
Channel
Manager
Sensor Node
y
i
(kk)
^
y
i
(kk1)
^
y
i
(kk) y
iq
(kkn)
^ ~
y
i
(kk) y
ip
(kkn)
^ ~
y
pi
(kk), y
qi
(kk)
^ ^
P
Q
Figure 61: Architecture of a decentralised data fusion node.
7.4.1 Preprocessing and Coordinate Transformations
When a sensor makes an observation, a number of preprocessing stages must occur to
transform observation data into information form (i(k) and I(k)) before it is fused in the
nodal lter.
The sensors employed are all range/bearing or bearing only. Observation information
is converted in to a Cartesian form for processing by the DDF algorithms. The Cosine
Matrix relating frame i to frame j (rotating in yaw, in pitch and in roll is
C
j
i
=
_
_
cos() cos() cos() sin() sin() sin() cos() cos() sin() cos() + sin() sin()
sin() cos() sin() sin() sin() + cos() cos() sin() sin() cos() cos() sin()
sin() cos() sin() cos() cos()
_
_
With this, the observed target is related to the sensor, the sensor to the body and then
the body to the earth frame, as,
P
e
t
= P
e
b
+C
e
b
P
b
s
+C
e
b
C
b
s
P
s
t
(365)
BAE SYSTEMS Proprietary Information, ASCTA use only 178
where
P
e
b
= [x
e
b
, y
e
b
, z
e
b
] is the position of the body in the earth frame and is provided by
the INS.
C
e
b
is the direction cosine matrix which relates how the body frame is rotated with
respect to the earth frame.
P
b
s
= [x
b
s
, y
b
s
, z
b
s
] is the position of the mission sensor to the body frame. This is
determined by calibration.
C
b
s
is the direction cosine matrix which relates how the sensor frame is rotated to
the body frame and is also provided by calibration. If the sensor frame is perfectly
aligned so that the frame axes of both the sensor and body frames are parallel then
the matrix is simply an Identity matrix.
P
s
t
= [x
s
t
, y
s
t
, z
s
t
] is the position of the target with respect to the sensor. The vector
which points to the target from the sensor frame can be described by two angles
and the range measurement. These angles will be taken as
s
t
which is the look
down angle of the target from the sensor frame on the xy plane. This angle will
transform the vector over to the xy plane from which the azimuth of the target
s
t
can be determined. Thus the position of the target becomes
P
s
t
=
_
_
x
s
t
y
s
t
z
s
t
_
_ =
_
_
r
s
t
cos
s
t
cos
s
t
r
s
t
cos
s
t
sin
s
t
r
s
t
sin
s
t
_
_ (366)
Note that for the radar/laser system the look down angle (grazing angle) is xed, that is,
s
t
is constant. Also note that positive looks down from the xy plane. The azimuth angle
is positive from the x axis clockwise when looking from above.
Once the measurements have been converted, it is necessary to convert the observation
noise to cartesian space. For a sensor with a standard deviation of
r
in range,
in
azimuth and
2
r
0 0
0 r
2
0
0 0 r
2
_
_
C
e
b
C
b
s
_
T
(367)
As the platform location is used in the observation coordinate transformation, it is neces
sary to adjust the observation information for the vehicle uncertainty. Failure to do this
could result in overly condent target estimates and possible lter divergence. This is
done using the following equation.
I
Norm
(k) = Y
T
Platform
(k [ k n)I(k)Y
Platform
(k [ k n) (368)
BAE SYSTEMS Proprietary Information, ASCTA use only 179
In summary, when a range/bearing observation is made at some time k, it is rstly
transformed to cartesian coordinates using Equation 365. The observation uncertainty
matrix is also transformed to the new reference frame by Equation 367. The observation
information matrix and information vector are then computed using Equation 209 and
the observation information matrix adjusted to account for the platform uncertainty by
Equation 368. This is necessary as the platform location information is used in the
coordinate transformation.
7.4.2 Data Association
Data association is necessary to correctly match information about the same target or
feature from dierent sources or at dierent time steps. When an observation is made, it
is necessary to determine if the target is same as one that has already been seen. Also,
in a decentralised system, it is necessary to associate information from other nodes with
that stored locally. Figure 62 illustrates this notion where decentralised system exists
with two nodes estimating the same targets, but they are ordered dierently. When node
one communicates information about its target 1, node 2 must correctly associate it with
its own target 3.
Target 1
Target 2
Target 3
Target 4
Target 5
Target n
Target 1
Target 2
Target 3
Target 4
Target 5
Target n
Targets at
Node 1
Targets at
Node 2
Figure 62: Dierent nodes may have the same physical targets stored in dierent orders.
The information gate and nearestneighbour algorithm is used for data association with
the information lter. This is very much a brute force approach. Nodes also include a
data association index with each information communication. The data association index
is the location of the target/feature at the transmitting node. When received for the
rst time, the targets pass through the data association algorithm to determine if they
match any targets at the receiving node. Once the receiving node knows the index of that
target locally, it can store the relationship between the targets on dierent nodes. In this
way, a look up table is generated once targets are identied. Figure 62 illustrates this.
BAE SYSTEMS Proprietary Information, ASCTA use only 180
A a lookup table mapping the targets on dierent nodes is shown. Along with the data
association index, nodes also include a probability of the data association index being
correct.
7.4.3 Local Filter
The local information lter generates information state estimates on the basis of observed,
predicted and communicated information. Other infrastructure such as the channel lter
and channel manager exist only to support the correct implementation of the local lter.
+ + Locall Filter
Local Filter
y
i
(kmkm)
Y
i
(kmkm)
y
i
(kkm)
Y
i
(kkm)
Coordinate
Transformation
Calc Obs
Information
Data
Association
Data
Association
Predict
i(k)
I(k)
z(k)
Z(k)
z(k)
Z(k)
y
i
(kk)
Y
i
(kk)
Sensor Preprocessing
i(k)
I(k)
i(k)
I(k)
Figure 63: Data ow from sensor observation.
The local lter takes input from local sensors (if present) and from the channel man
ager (if connected). Local sensors preprocess observation data to produce an observation
information vector i(k), observation information matrix I(k), a data association index,
and an associated probability of correctness to the lter. The generation of observations is
typically asynchronous. This observation information is communicated to the local node
lter where (see Figure 63:
1. The local information state is predicted forward to observation time.
2. It passes through a data association stage using the a combination of both the
information gate and the data association index sent with the information package.
3. With a correct association, the node lter then fuses observation and prediction
information through the summation in Equation 321.
4. The fused estimate is then propagated forward from observation time to the syn
chronous node step.
5. Equation 312 is generally used to predict the information state forward in time. If
data is delayed in communication or observation to the node lter, it is fused using
delayed data Equation 317 given in Section 5.2.8.
At the given node synchronous rate, the node lter also receives new information from
the channel manager. The state is predicted forward to this time using Equation 312 and
updated with Equation 321. The full information matrix and information vector are then
output to the channel manager for transmission to neighbouring nodes.
BAE SYSTEMS Proprietary Information, ASCTA use only 181
7.4.4 Channel Filter
The channel lter is used to manage communication between nodes. It serves two main
functions; to keep track of information previously communicated on the channel, and to
synchronize incoming and outgoing information with the node lter. Information previ
ously communicated is used to compute new information gain to other sensor nodes in the
network. Synchronization serves to accommodate any delays in information or dierences
in timing between node lters at remote sites.
Figure 64 shows the ow of incoming information through the channels, channel man
ager and local lter. Information arrives asynchronously at each channel from remote
nodes. The information rst passes through a preprocessing stage where tracktotrack
association is performed using both data association index and an information gate, in or
der to associate the new data with an existing lter track. The information state is then
predicted forward to the time horizon using the standard prediction method of Equa
tion 312. As the data enters the channel lter, the new information is determined and is
then transmitted to the channel manager. The channel lter is then updated.
Buffer Data Association Predict Buffer +  Predict Channel Filter
y
a
(kkn
a
)
Y
a
(kkn
a
)
Channel Filter Preprocess
y
ia
(kmkm)
Y
ia
(kmkm)
y
ia
(kkm)
Y
ia
(kkm)
y
b
(kn
b
kn
b
)
Y
b
(kn
b
kn
b
)
Buffer Data Association Predict Buffer +  Predict Channel Filter
y
b
(kkn
b
)
Y
b
(kkn
b
)
Channel Filter Preprocess
y
ib
(kmkm)
Y
ib
(kmkm)
y
ib
(kkm)
Y
ib
(kkm)
Buffer Data Association Predict Buffer +  Predict Channel Filter
y
c
(kkn
c
)
Y
c
(kkn
c
)
Channel Filter Preprocess
y
ic
(kmkm)
Y
ic
(kmkm)
y
ic
(kkm)
Y
ic
(kkm)
Channel Manager + + Predict Local Filter
Local Filter
y
i
(kmkm)
Y
i
(kmkm)
y
i
(kkm)
Y
i
(kkm)
y
c
(kn
c
kn
c
)
Y
c
(kn
c
kn
c
)
y
a
(kn
a
kn
a
)
Y
a
(kn
a
kn
a
)
Channel a
Channel b
Channel c
y
c/i
(k)
Y
c/i
(k)
y
b/i
(k)
Y
b/i
(k)
y
a/i
(k)
Y
a/i
(k)
Figure 64: Flow of information for incoming channel data.
Figure 65 shows the ow of outgoing information from the nodal lter to the channel
manager and to the channels. As information is sent to the channel, the channel lter
is updated and the current state is transmitted down the channel. In the event that the
channel becomes blocked or disconnected, the channel lter eectively fuses the new data
and cycles to the next available communication time.
7.4.5 Channel Manager
The channel manager serves as the interface between the nodal lter and the channel
lters (and through these, the other nodes in the network). The channel manager collects
incoming data from the channels at the time horizon, assimilates it, and then communi
cates the result to the nodal lter. It also receives outgoing updated information states
BAE SYSTEMS Proprietary Information, ASCTA use only 182
y
i
(kk)
Y
i
(kk)
y
i
(kk)
Y
i
(kk)
y
i
(kk)
Y
i
(kk)
Channel Filter
y
i
(kk)
Y
i
(kk)
y
i
(kk)
Y
i
(kk)
Channel Filter
y
i
(kk)
Y
i
(kk)
y
i
(kk)
Y
i
(kk)
Decentralised Node
Channel Filter
Channel
Manager
Local Filter
Figure 65: Flow of information for outgoing channel data.
from the nodal lter and disseminates this to the channel lters for transmission. The
channel manager manages online channel connectivity and allocation of, the allocation is
handled by the channel manager as well.
7.4.6 Timing
Each clock at each node is synchronised to a common system time, in order to eliminate
errors in clock drifts.
Every node in the decentralised system operates asynchronously with respect to the
other nodes. The nodes are programmed to communicate their information every time
period T (where T is the period between time horizons), which need not be the same
for all nodes. When a node receives new information through a channel, it is predicted
to the time horizon of the local node and temporarily buered. The time horizon is a
time in the future at which the local node will assimilate the new data from all of its
channels with the local state estimate. The prediction to this horizon is done using the
same model as the local lter. Thus, data will arrive asynchronously at a nodes channel
lter, then be predicted and stored until the next time horizon. Figure 66 illustrates this
timing. If a situation arises where two pieces of information arrive on the same channel
between any two time horizons, only the most recent to arrive should be used and the
other(s) discarded. As the channels have a one to one mapping, the most recent of these
measurements will always contain all the information of the earlier message in addition
to any newer data. This situation can occur if dierent nodes are transmitting their
BAE SYSTEMS Proprietary Information, ASCTA use only 183
information at dierent frequencies, a scenario which is entirely valid under the system
specications. However, if the channel lter is implemented correctly then this will not
cause a problem.
Time
Time
Horizon
t(k1)
Channel 1
Channel 2
Channel n
Time
Horizon
t(k)
Time
Horizon
t(k+1)
Data arrives through
Channel 1 and is
predicted forward to
horizon at t(k1)
At time horizon all
channels are checked
and new data is fused
in filter.
Figure 66: Data arrives at the channel lter asynchronously from other nodes. It is
predicted forward to a local time horizon and fused then.
It is important that data not be transmitted from the node if there is still new data
buered in the channels. Prior to transmitting information, the channels should be read
and cleared. This is necessary as the channel lter at the transmitting node will already
have updated its channel lter with the data waiting in the buer at the receiving node,
and if that receiving node transmits information back that does not include this infor
mation then the system becomes inconsistent. In practice, what occurs is at the time
horizon the channels and the local lter are updated and the node then outputs its new
state then. That is, each node outputs it state at the local time horizon after it has
updated with the information in the channels. In this way, the timing within each node is
kept synchronous, but the timing between nodes is asynchronous. This ensures the lters
remain consistent and do not have any timing ambiguities.
If the system is implemented in this manner, it also removes any delayed or asequent
data problem through the channel lter. When the data arrives, it is automatically
predicted forward to the time horizon. There still exists the possibility of delayed data
from the local sensor, however, this can be handled using the delayed data algorithm of
Section 5.2.8.
7.4.7 Picture Compilation
For the decentralised picture compilation problem, a separate channel lter/local lter
pair is maintained for each target. It is possible to augment all of the targets into a single
BAE SYSTEMS Proprietary Information, ASCTA use only 184
vector, however as the targets are uncorrelated the information matrix is sparse.
In picture compilation mode, the location of the sensors and platform is assumed
known with precision. This location information is used to convert all relative observation
information to a global cartesian ground coordinates in which tracking takes place. This
conversion is accomplished in the sensor preprocessing stage for both bearing only and
range and bearing sensor types (see Section 7.4.1).
In global ground coordinates, all targets are modeled using a threedimensional con
stant velocity ballistic model in the form:
x(k) = F
(k,k1)
x(k 1 [ k 1) +w
(k,k1)
(369)
where the state vector is
x(k) = [x(k), x(k), y(k), y(k), z(k), z(k)]
T
(370)
The state transition matrix for this system is given by
F
(k,k1)
=
_
_
1 T 0 0 0 0
0 1 0 0 0 0
0 0 1 T 0 0
0 0 0 1 0 0
0 0 0 0 1 T
0 0 0 0 0 1
_
_
(371)
The process noise is written as G
k
Q
(k,k1)
G
T
k
where
Q
(k,k1)
=
_
_
q
x
0 0
0 q
y
0
0 0 q
z
_
_ (372)
and
G
k
=
_
_
T
2
2
0 0
T 0 0
0
T
2
2
0
0 T 0
0 0
T
2
2
0 0 T
_
_
(373)
While the real observations are of the range/bearing type, they are converted to cartesian
coordinates in order to linearise the problem. After this conversion has been made, the
observation matrix is simply given by
H(k) =
_
_
1 0 0 0 0 0
0 0 1 0 0 0
0 0 0 0 1 0
_
_ (374)
where the observations are in the form z(k) = [x, y, z]
T
.
BAE SYSTEMS Proprietary Information, ASCTA use only 185
The picture compilation system uses an information gate for data association. Nodes
also use the data association index to map targets at one node to targets at another, as
detailed in Section 7.4.2.
The coordinate transformations to convert the range/bearing measurements to carte
sian will be done in the sensor preprocessing stage. The equations, given in Section 7.4.1,
convert the measurement and the observation covariance taking into account the platform
uncertainty. Once the sensor preprocessing has completed these transformations, the ob
servation information vector is calculated according to Equation 209 and transmitted to
the decentralised lter node.
7.4.8 Terrain Navigation
The decentralised terrain navigation system is a single vehicle problem. Each platform
has all of its sensors decentralised and communicating internally, however, there is no
interplatform communication. The platforms start with a prior map and navigate by
observing these known features. Each node in the system runs a single local lter (and
associated channel manager and channel lters) which estimates the platform position,
velocity and attitude. The state vector for this is given by
x(k) = [x(k), y(k), z(k), x(k), y(k), z(k), (k), (k), (k)]
T
(375)
The decentralised system includes some combination of radar, laser and camera as
well as an IMU and GPS. The local lter at the inertial node will run a lter estimating
the same states as the rest of the system, but need not necessarily be in the same linear
form as other nodes. The result of this is that when the inertial node outputs its state
through the channel lter, it may in eect be outputting a loss in information. When
the node receiving this information determines the information gain, it will be therefore
be negative. The prediction stage of the channel lter in the inertial node uses the same
linear model as other nodes for the time alignment of data.
The extension to the Simultaneous Localisation and Mapping (SLAM) problem in
volves an expanded state incorporating the locations of all (stationary) land marks.
7.4.9 The Communication Layer
While the decentralised algorithms are to be implemented in a point to point network, the
physical communication medium onboard the Brumby platforms is a CAN bus. Therefore,
the point to point architecture will be developed in software using virtual channels to
connect nodes. This is illustrated in Figure 67 where multiple decentralised nodes can be
seen to reside on a single processor.
The virtual channels are supported by the CommLib communication software devel
oped for the project. In addition to the connections between the nodes, separate channels
are required for
DDF links between node pairs
BAE SYSTEMS Proprietary Information, ASCTA use only 186
CAN Bus
DDF
Node
DDF
Node
DDF
Node
Radar
Camera
PC/104
PC/104
Figure 67: A virtual tree network connected on a CAN bus.
The platform states from the GPS/IMU process
Time synchronisation messages
All nodes in the system will be synchronised with the ight control computer, which
will be synchronised with GPS time in order to keep the clocks across dierent platforms
from drifting. If the GPS signal is lost for a period of time, all nodes within a platform
stay synchronised with the ight control computer, however, the ight control computers
on dierent platforms may drift as they are not being reset.
7.5 Large Scale Sensor Networks
Work is also underway at Sydney to develop a groundbased sensor network to be located
at the same test site as the ight platforms. The aim of this work is to demonstrate large
scale scalability of decentralised data fusion methods and to show distribution across both
air and land environments.
Currently the project is developing a group of groundbased sensor nodes, based on the
same architecture as the ANSER project. The nodes include vision, and multispectral
cameras, lasers and (in the future) radars. Sensor nodes are being physically constructed
in a modular fashion around a PC104 architecture and linked by radio ethernet.
The network will be used for both airtarget tracking and hando of ground targets
from air to land sensors. The practical objectives are to demonstrate the applicability of
decentralised methods to a broader range of data fusion problems, and to serve as the
primary testvehicle for systemofsystem theories and algorithms.
The largescale sensor network laboratory associated with this course aims to explore
some of the issues in largescale sensor networks. In particular, the issue of dynamic
BAE SYSTEMS Proprietary Information, ASCTA use only 187
communication topologies, and network organisation. Figure 68 shows the typical network
congurations to be explored. Figure 69 shows typical tracks. Figure 70 shows tracking
results from a typical node.
BAE SYSTEMS Proprietary Information, ASCTA use only 188
3 4 5 6 7 8
x 10
4
2
3
4
5
6
7
8
x 10
4
Platforms and Communication Links
xposition (m)
y
p
o
s
it
io
n
(
m
)
Sensors
Communication Links
(a)
2 3 4 5 6 7 8
x 10
4
2
3
4
5
6
7
8
x 10
4
Platforms and Communication Links
xposition (m)
y
p
o
s
it
io
n
(
m
)
Sensors
Communication Links
(b)
3 4 5 6 7 8
x 10
4
2
3
4
5
6
7
8
x 10
4
Platforms and Communication Links
xposition (m)
y
p
o
s
it
io
n
(
m
)
Sensors
Communication Links
(c)
2 3 4 5 6 7 8
x 10
4
1
2
3
4
5
6
7
8
x 10
4
Platforms and Communication Links
xposition (m)
y
p
o
s
it
io
n
(
m
)
Sensors
Communication Links
(d)
Figure 68: Dierent communication topologies for a 20 node network: (a) Fully connected;
(b) and (c) tree connected; (d) general network topology.
BAE SYSTEMS Proprietary Information, ASCTA use only 189
0 1 2 3 4 5 6 7 8 9 10
x 10
4
0
1
2
3
4
5
6
7
8
9
10
x 10
4
True Target Motions
xposition (m)
y
p
o
s
i
t
i
o
n
(
m
)
Target True Position
Tracking Stations
Target Observations
(a)
2.6 2.8 3 3.2 3.4 3.6 3.8
x 10
4
7.4
7.6
7.8
8
8.2
8.4
8.6
8.8
x 10
4
Estimated Target Motions
xposition (m)
y
p
o
s
i
t
i
o
n
(
m
)
(b)
Figure 69: Typical tracks and observations generated: (a) true xy target tracks and
observations; (b) detail of estimated tracks from all nodes.
BAE SYSTEMS Proprietary Information, ASCTA use only 190
0 50 100 150 200 250 300 350 400 450 500
300
250
200
150
100
50
0
50
100
150
200
Node 1: X position Error and Covariance
time
X
e
r
r
o
r
(
m
)
Actual Error
Estimated Error
(a)
0 50 100 150 200 250 300 350 400 450 500
80
60
40
20
0
20
40
60
80
Node 1: Y position Error and Covariance
time
Y
e
r
r
o
r
(
m
)
Actual Error
Estimated Error
(b)
0 50 100 150 200 250 300 350 400 450 500
80
60
40
20
0
20
40
60
80
Node 1: Y position Error and Covariance
time
Y
e
r
r
o
r
(
m
)
Actual Error
Estimated Error
(c)
0 50 100 150 200 250 300 350 400 450 500
0.2
0.1
0
0.1
0.2
0.3
0.4
0.5
Node 1: Heading Error
time
H
e
a
d
in
g
e
r
r
o
r
(
r
a
d
s
)
(d)
Figure 70: Estimates from a node: (a) xposition; (b) yposition; (c) velocity; (d) heading.
BAE SYSTEMS Proprietary Information, ASCTA use only 191
References
[1] B.D.O. Anderson and J.B. Moore. Optimal Filtering. Prentice Hall, 1979.
[2] Y. BarShalom. On the track to track correlation problem. IEEE Trans. Automatic Control,
25(8):802807, 1981.
[3] Y. BarShalom. MultiTarget MultiSensor Tracking. Artec House, 1990.
[4] Y. BarShalom. MultiTarget MultiSensor Tracking II. Artec House, 1990.
[5] Y. BarShalom and T.E. Fortmann. Tracking and Data Association. Academic Press, 1988.
[6] T. Berg. Model Distribution in Decentralized MultiSensor Data Fusion. PhD thesis, The
University of Oxford, 1993.
[7] T. Berg and H. DurrantWhyte. On distributed and decentralized estimation. In American
Control Conference, pages 22734, 1992.
[8] T. Berg and H.F. DurrantWhyte. Model distribution in decentralized multisensor fusion.
In Proc. American Control Conference, pages 22922294, 1991.
[9] T. Berg and H.F. DurrantWhyte. Model distribution in decentralized sensor networks. In
IEEE Int. Conf. on Intelligent Control and Instrumentation, 1992.
[10] J.O. Berger. Statistical Decision Theory and Bayesian Analysis. Springer Verlag, 1985.
[11] S. Blackman and R. Popoli. Design and Analysis of Modern Tracking Systems. Artec
House, 1999.
[12] R.G. Brown and P.Y.C. Hwang. Introduction to Applied Kalman Filtering. John Wiley,
1992.
[13] T. Burke. Design and Control of a Modular Wheeled Mobile Robot. PhD thesis, The
University of Oxford, 1994.
[14] T. Burke and H.F. DurrantWhyte. Modular mobile robot design. In 1st IFAC International
Workshop on Intelligent Autonomous Vehicles, Southampton, U.K., 1993.
[15] D.E. Catlin. Estimation, Control and the Discrete Kalman Filter. Springer Verlag, 1984.
[16] C. Chong, K. Chang, and Y. BarShalom. Joint probabilistic data association in distributed
sensor networks. IEEE Trans. Automatic Control, 31(10):889897, 1986.
[17] T.M. Cover and J.A. Thomas. Elements of Information Theory. John Wiley, 1991.
[18] R.H. Deaves. The Management of Communications in Decentralised Bayesian Data Fusion
Systems. PhD thesis, The University of Bristol, 1998.
[19] R.H. Deaves, P. Greenway, and D.R. Bull. Communications management in decentralised
data fusion systems. In IEEE In. Conf. on Multisensor Fusion and Integration for Intelligent
Systems, 1996.
BAE SYSTEMS Proprietary Information, ASCTA use only 192
[20] H.F. DurrantWhyte. Introduction to Estimation and The Kalman Filter. Australian Centre
for Field Robotics, 2000.
[21] M. Fernandez. Failure Detection and Isolation in Decentralised Multisensor Systems. PhD
thesis, The University of Oxford, 1994.
[22] M. Fernandez and H. DurrantWhyte. An informationtheoretic approach to fault detection
and identication. In American Control Conference, 1993.
[23] M. Fernandez and H. DurrantWhyte. A failure detection and isolation algorithm for de
centralized multisensor systems. In IEEE Int. Conf. on Multisensor Fusion and Integration
for Intelligent Systems, 1994.
[24] S. Grime. Communication in Decentralised Sensing Architectures. PhD thesis, The Uni
versity of Oxford, 1992.
[25] S. Grime and H. DurrantWhyte. Communication in decentralized systems. IFAC Control
Engineering Practice, 2(5):849863, 1994.
[26] S.Y. Harmon. Sensor data fusion through a blackboard. In Proc. IEEE Int. Conf. Robotics
and Automation, page 1449, 1986.
[27] C.J. Harris and I. White. Advances in Command, Control and Communication Systems.
IEE press, 1987.
[28] H.R. Hashemipour, S. Roy, and A.J. Laub. Decentralized structures for parallel Kalman
ltering. IEEE Trans. Automatic Control, 33(1):8893, 1988.
[29] P. Ho. Organisation in Decentralised Sensing. PhD thesis, The University of Oxford, 1995.
[30] T. Kilath. Linear Systems. John Wiley, 1980.
[31] P. Lancaster and M. Tismenetsky. The Theory of Matrices. Academic Press, 1985.
[32] S.M. Magrabi and P.W. Gibbens. Decentralised fault detection and diagnosis in guidance
and navigation systems. In IEEE Position Location and Navigation Symposium, 2000.
[33] J. Manyika. An Information Theoretic Approach to Data Fusion and Sensor Management.
PhD thesis, The University of Oxford, 1993.
[34] J. Manyika and H. DurrantWhyte. Decentralized sensor management. In Proc. SPIE Conf.
Data Fusion, 1992.
[35] J. Manyika and H. DurrantWhyte. Information as a basis for management and control
in decentralized fusion architectures. In Proc. IEEE Conf. Decision and Control, Tuscon,
1992.
[36] J. Manyika and H.F. DurrantWhyte. Data Fusion and Sensor Management: An
InformationTheoretic Approach. Prentice Hall, 1994.
[37] P.S. Maybeck. Stochastic Models, Estimaton and Control, Vol. I. Academic Press, 1979.
BAE SYSTEMS Proprietary Information, ASCTA use only 193
[38] A.G.O Mutambara and H.F. DurrantWhyte. The control of a modular mobile robot. In
IEEE Conf. Intelligent Robotics (IROS), 1994.
[39] A.G.O Mutambara and H.F. DurrantWhyte. Modular decentralized control. In American
Control Conference, 1994.
[40] A.G.O Mutambara and H.F. DurrantWhyte. Nonlinear information space: a practical
basis for decentralization. In Proc.SPIE 94 Conference: Sensor Fusion VII, volume 2355,
1994.
[41] A.G.O Mutambura. Decentralised Control. PhD thesis, The University of Oxford, 1995.
[42] A.G.O Mutambura. Decentralised Estimation and Control for Multisensor Systems. CRC
Press, 1998.
[43] J.F. Nash. The bargaining problem. Econometrica, page 155, 1950.
[44] D. Nicholson, S. Julier, and J. Uhlmann. Ddf an evaluation of covariance intersection. In
Fusion 2001, 2001.
[45] H.P. Nii. Blackboard systems. AI Magazine, 1986.
[46] A. Papoulis. Probability, Random Variables, and Stochastic Processes; Third Edition.
McGrawHill, 1991.
[47] B. Rao and H. DurrantWhyte. A decentralized bayesian algorithm for identication of
tracked targets. IEEE Trans. Systems Man and Cybernetics, 23(6):16831698, 1993.
[48] B. Rao, H. DurrantWhyte, and A. Sheen. A fully decentralized multisensor system for
tracking and surveillance. Int. J. Robotics Research, 12(1):2045, 1991.
[49] B.S.Y. Rao. Data Fusion Methods in Decentralised Sensing Systems. PhD thesis, The
University of Oxford, 1991.
[50] B.S.Y. Rao and H.F. DurrantWhyte. A fully decentralized algorithm for multisensor
kalman ltering. IEE Transactions Schedule D, 138(5):413420, 1991.
[51] C.R. Rao. Linear Statistical Inference and its Applications. John Wiley, 1965.
[52] N.R. Sandell, P. Varaiya, M. Athans, and M.G. Safonov. Survey of decentralized control
methods for large scale systems. IEEE Trans. Automatic Control, 23(2):108128, 1978.
[53] H.W. Sorenson. Special issue on applications of Kalman ltering. IEEE Trans. Automatic
Control, 28(3), 1983.
[54] J.L. Speyer. Communication and transmission requirments for a decentralized linear
quadraticgaussian control problem. IEEE Trans. Automatic Control, 24(2):266269, 1979.
[55] L.D. Stone, C.A. Barlow, and T.L. Corwin. Bayesian Multiple Target Tracking. Artech
House, 1999.
[56] G. Strang. Linear Algebra and its Applications, Third Edition. Harcourt Brace Jovanovich,
1988.
BAE SYSTEMS Proprietary Information, ASCTA use only 194
[57] S. Utete. A network manager for a decentralized sesning system. Masters thesis, University
of Oxford, 1992.
[58] S. Utete. Network Management in Decentralised Sensing Systems. PhD thesis, The Uni
versity of Oxford, 1994.
[59] S. Utete and H.F. DurrantWhyte. Management for reliability of decentralised sensing
networks. In Proc. 1994 World Transputer Congress, IOS Press, 1994.
[60] S. Utete and H.F. DurrantWhyte. Network management in decentralised data fusion
networks. In Proc. SPIE 94 Conference, Sensor Fusion VII, volume 2355, 1994.
[61] S. Utete and H.F. DurrantWhyte. Reliability in decentralised data fusion networks. In
IEEE Int. Conf. on Multisensor Fusion and Integration for Intelligent Systems, pages 215
221, 1994.
[62] S. Utete and H.F. DurrantWhyte. Routing for reliability in decentralized sensing networks.
In American Control Conference, 1994.
[63] E.L. Waltz and J. Llinas. Sensor Fusion. Artec House, 1991.
Much more than documents.
Discover everything Scribd has to offer, including books and audiobooks from major publishers.
Cancel anytime.